Robot Wiki

Credits

Every photograph and diagram on Robot Wiki is listed here with its creator, source page and licence. Article images are CC0, CC BY, CC BY-SA, public domain or used under a documented reuse permission. No image is AI-generated.

Market-map logos include official company marks, recorded as unlicensed or unknown where no reuse grant is named. The site's own text and original diagrams are licensed CC BY 4.0.

Written and maintained by Josef Chen (@josefchen).

Who is behind Robot Wiki

I spent three years building and deploying robots at KAIKAKU (acquired by REEF). What I needed over those years was scattered across papers that skip the deployment context and demos that never quote a failure rate. Robot Wiki collects it for engineers who come from machine learning and cites the source of every claim.

Corrections follow the editorial policy; report a factual issue on GitHub.

  1. The 2013 DARPA Atlas humanoid robot stands facing forward with its exposed metal frame, hydraulic lines, and stereo camera head visible.
    The original Atlas humanoid, built for the 2013 DARPA Robotics Challenge: hydraulically actuated, tethered, with its frame and sensor head exposed. The electric Atlas discussed above is its clean-sheet successor.
    Photo: DARPA / Wikimedia Commons. Licence: Public domain.

    Appears on: Hardware Taxonomy

  2. A stroboscopic NASA photograph shows a PUMA 560 robot arm in several successive positions as it moves small parts between fixtures on a table.
    A PUMA 560 at NASA Ames, photographed in a stroboscopic exposure as it cycles through its workspace. The six-revolute PUMA layout is the canonical teaching example for the Denavit-Hartenberg convention.
    Photo: NASA / Dominic Hart / Wikimedia Commons. Licence: Public domain.

    Appears on: Kinematics

  3. A red and grey ANYmal quadruped robot stands on a metal grate walkway inside an industrial facility, surrounded by pipes and machinery.
    ANYmal, the ANYbotics quadruped behind the ETH Zurich locomotion lineage, on an industrial inspection floor. The actuator-network and terrain-curriculum results above all ran on this platform.
    Photo: ANYbotics / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Legged Locomotion Lineage

  4. A white seven-axis Franka Emika Panda robot arm mounted on a demonstration table at CeBIT 2017, guided by hand above trays of small parts.
    The Franka Emika Panda, the seven-DoF research arm, demonstrated at CeBIT 2017 with a human guiding it by hand.
    Photo: Ims / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Hardware Taxonomy

  5. Covariate shift
    One small slip, and the robot leaves familiar ground
    Original schematicpaths a person showed itthe robot aloneFirst slip: from here onit has no examples to copy
    Robots trained by copying only learn the situations they were shown; one slip lands them somewhere new, where slips grow.
    How this was made

    A schematic, not data. A robot trained by behavior cloning learns from the states an expert visited (the shaded corridor of demonstrations). Running on its own, it visits its own states, and the two drift apart after its first mistake: from there it is outside anything it was trained on, so its next mistake is likely to be larger.

    The mismatch between the training states and the states the robot reaches is called covariate shift, or distribution shift. Ross, Gordon and Bagnell (2011) showed that the resulting cost can grow with the square of the task length, and proposed DAgger, which has an expert label the states the robot actually visits.

    Diagram: Robot Wiki contributors / Robot Wiki (original diagram). Licence: CC BY 4.0.

    Appears on: Behavior Cloning Foundations

  6. Temporal ensembling
    Three overlapping plans vote; the oldest gets the loudest voice
    Original schematicoldest plan, 51%next plan, 31%newest plan, 19%the move it makes
    Blending several overlapping plans smooths the robot’s motion, so one odd prediction jerks the arm less.
    How this was made

    A schematic, not data. In ACT the robot asks its policy for a fresh chunk of future moves at every control step, so the chunks issued at earlier steps each still hold a prediction for the current move. Temporal ensembling averages those predictions, and the oldest gets the largest weight.

    Counting from the oldest (i = 0), prediction i gets the raw weight w_i = exp(−m·i); the weights are then divided by their sum. The drawing uses m = 0.5 so the three weights differ visibly: raw 1.00, 0.61, 0.37, which is 51%, 31%, 19% of the vote. The ACT reference code uses m = 0.01, which makes the weights almost equal.

    Diagram: Robot Wiki contributors / Robot Wiki (original diagram). Licence: CC BY 4.0.

    Appears on: Action Chunking (ACT and ALOHA)

  7. A rendered ALOHA 2 workcell: two pairs of black robot arms facing each other across a table inside an aluminium frame with overhead cameras, with letters A, B and C marking measured segments.
    ACT was introduced with ALOHA, a low-cost bimanual rig on which a person demonstrates by moving leader arms. A render of the ALOHA 2 workcell, with its frame segments labelled A, B and C.
    Paper figure: Tony Z. Zhao et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Action Chunking (ACT and ALOHA)

  8. Five photographs of ALOHA 2 bimanual robot arms at work, labelled Robot Finger Replace, Shirt Hanging, Gear Insertion, Shoelace Tying and Kitchen Stacking.
    Chunked policies now do dexterous bimanual work such as tying shoelaces and hanging shirts. ALOHA Unleashed collected over 26,000 demonstrations for these five tasks on ALOHA 2 and ran each predicted chunk open loop.
    Paper figure: Tony Z. Zhao et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Action Chunking (ACT and ALOHA)

  9. Photographs of two robot platforms, labelled AgileX PiPER Single/Dual Arm Platform and AIRBOT Platform, with arrows marking the head and wrist camera views.
    The two arm platforms of an action-space study with more than 13,000 real rollouts and 500+ models. Predicting delta actions consistently helped; joint space favoured control stability and task space favoured generalisation.
    Paper figure: Yuchun Feng et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Action Spaces for Robot Learning

  10. Two photographs, each beside a line drawing: a robot arm pushes a blue box up against a block, with arrows marking a high stiffness K1 and a low stiffness K2, and two arms hold and wipe a white vase, with arrows marking compliance directions.
    An impedance action sets the robot's stiffness in each direction as well as its goal. Adaptive Compliance Policy flips an item, soft (K2) along the push and stiff (K1) elsewhere, and wipes a vase with both arms compliant.
    Paper figure: Yifan Hou et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Action Spaces for Robot Learning

  11. A labelled photograph of the DROID data collection rig: a Franka Panda arm on a height-adjustable desk with two side-mounted Zed 2 cameras, a wrist camera and an Oculus Quest 2 headset with controllers.
    Behavior cloning starts from demonstrations recorded on rigs like this one, where a person teleoperates the arm through a VR headset. DROID used this Franka Panda setup at every institution.
    Paper figure: Alexander Khazatsky et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Behavior Cloning Foundations

  12. Nine photographs of robot setups: WidowX BridgeV2, UR5 Tabletop and RT-1 Robot grouped under 0-Shot Evaluation, and Berkeley Insert, Stanford Coffee, CMU Baking, Berkeley Pick-up, Berkeley Coke and Berkeley Bimanual grouped under Finetuning Evaluation.
    Octo, an open-weights policy, was evaluated out of the box on three robot setups and fine-tuned on six more, some with new observations, action spaces or embodiments.
    Paper figure: Octo Model Team / arXiv. Licence: CC BY 4.0.

    Appears on: Comparison Matrix, Vision-Language-Action Models

  13. A diagram of RoboArena: a pool of generalist policies feeds a ring of photographed robot workstations, each running pairwise comparisons on its own task, which aggregate into a ranked score table.
    RoboArena ranks policies from double-blind pairwise trials that many evaluators run on their own scenes and tasks. More than 600 real-robot episodes on the DROID platform ranked seven generalist policies.
    Paper figure: Pranav Atreya et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Evaluation Crisis, Comparison Matrix, Other Generalist Policies, World Model Evaluation for Robotics

  14. Two rows of frames: an Apptronik Apollo humanoid robot packs a lunch bag, and a bi-arm industrial Franka setup fits a rubber band around a pulley system.
    One Gemini Robotics model was fine-tuned to drive very different bodies, from the Apollo humanoid packing a lunch bag (top) to a bi-arm Franka fitting a rubber band around pulleys (bottom).
    Paper figure: Gemini Robotics Team / arXiv. Licence: CC BY 4.0.

    Appears on: Comparison Matrix, Other Generalist Policies

  15. An infographic framed by robot dataset thumbnails reading 1M episodes from 311 scenes, 34 research labs across 21 institutions, 22 embodiments, 527 skills and 60 datasets.
    Open X-Embodiment pooled robot data into one training set: 1M episodes from 60 datasets, 22 embodiments and 527 skills, contributed by 34 labs at 21 institutions.
    Paper figure: Open X-Embodiment Collaboration / arXiv. Licence: CC BY 4.0.

    Appears on: Major Datasets, Cross-Embodiment Transfer, Vision-Language-Action Models

  16. A three-stage diagram with photographs: egocentric human videos of hands doing chores, paired human and robot play data, and a humanoid robot hand performing dexterous tasks after post-training.
    EgoScale pretrains on human video, mid-trains on aligned human and robot play, then post-trains on robot tasks. The pretraining stage uses 20,854 hours of egocentric video with wrist motion and retargeted hand actions.
    Paper figure: Ruijie Zheng et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Cross-Embodiment Transfer

  17. A pipeline of photographs: a human hand demonstrates a task, the hand is inpainted out of the video, a rendered robot gripper is overlaid in its place, and a real robot then performs the task.
    Phantom trains a robot policy without robot data by erasing the human hand from demonstration videos and rendering a robot arm in its pose. The policy then deploys zero-shot on a real robot, with success up to 92%.
    Paper figure: Marion Lepert et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Cross-Embodiment Transfer

  18. Six labelled frames of two robot arms hanging a red shirt: initial state, orient shirt, pick hanger, insert right collar, insert left collar and hang on rack.
    A diffusion policy hangs a shirt on its own, a long-horizon task with a deformable object and several bimanual stages. Six frames of an autonomous ALOHA Unleashed rollout, from the initial state to the shirt on the rack.
    Paper figure: Tony Z. Zhao et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Diffusion Policy

  19. Two photographs: everyday objects on a wooden table with the letters DP3 spelled out, and a Franka arm with an Allegro hand and a RealSense L515 camera above a table.
    3D Diffusion Policy swaps images for sparse point clouds from a single depth camera. Left, the everyday objects in its real tasks; right, a Franka arm with an Allegro hand under a RealSense L515.
    Paper figure: Yanjie Ze et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Diffusion Policy

  20. Photographs of the robot arms RoboCat controls, each labelled with its action dimensions: Sawyer and Panda arms with parallel grippers seen in training, and a KUKA arm with a three-finger hand held out for fine-tuning.
    An adaptation budget is part of a foundation claim: RoboCat learned Sawyer and Panda arms in training and met the KUKA arm with a three-finger hand only at fine-tuning, from 100 to 1,000 examples.
    Paper figure: Konstantinos Bousmalis et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Foundation Models for Robotics

  21. Rows of frames from a Fourier GR-1 humanoid robot at a wooden table handing objects between its hands, placing a peach in a bin and moving items onto plates and into baskets.
    GR00T N1 runs a vision-language model at 10 Hz over a 120 Hz action module on the GR-1 humanoid. Frames from real rollouts: pretraining evaluations (top) and post-trained pick-and-place and articulated-object tasks (bottom).
    Paper figure: NVIDIA et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Foundation Models for Robotics, Hierarchical Approaches

  22. Two rows of ALOHA robot frames with arrows between them: putting an object on a plate, folding a shirt, putting candies in a bowl and putting candy in a zip bag.
    Cosmos Policy fine-tunes a pretrained video model into a robot policy, so a predictive model becomes the controller. Rollouts on four ALOHA tasks, from folding a shirt to putting candy in a zip bag.
    Paper figure: Moo Jin Kim et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Foundation Models for Robotics, What Is a World Model?

  23. Four rows of overhead frames of two robot grippers opening an eyeglasses case, pouring pulses, unfastening a file folder and wrapping a headphone wire on a cluttered table.
    A vision-language-action model performs dexterous bimanual work in clutter. Gemini Robotics on ALOHA, top to bottom: open the eyeglasses case, pour pulses, unfasten a file folder and wrap a headphone wire.
    Paper figure: Gemini Robotics Team / arXiv. Licence: CC BY 4.0.

    Appears on: Other Generalist Policies, Vision-Language-Action Models

  24. Two rows of frames of a robot arm pouring from a paper cup into a bowl and wiping a table with a sponge, each labelled with language motions such as rotate arm left and move arm forward and left.
    RT-H first predicts a language motion such as "rotate arm left", then the action, so a person can correct the motion in words. Frames from two unseen tasks, with and without such corrections.
    Paper figure: Suneel Belkhale et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Hierarchical Approaches

  25. Two pairs of photographs of real robot tabletop scenes, each with a coloured path drawn from the gripper towards its target: screwing a light bulb into a lamp, and pushing coloured buttons in order.
    HAMSTER's high level draws a coarse 2D path for the gripper, and a low-level policy turns the path into motion. Paths predicted on real scenes run from blue to red; blue and red circles mark grasp and release.
    Paper figure: Yi Li et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Hierarchical Approaches

  26. A strip of frames from a robot camera showing a mobile robot clearing dishes, cups and trash from a table into a bin.
    Table bussing is the task behind KI's training-cost result: a many-embodiment π0 generalist needed 7.5 times the training steps of π0.5 with KI to reach similar performance. Frames of the task from the FAST paper.
    Paper figure: Karl Pertsch et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Knowledge Insulation

  27. An overhead view of two robot arms folding a flat cardboard sheet into a box on a table.
    π0.6, whose card lists a Gemma3 4B backbone with an insulated expert of about 860M parameters, building a box. Physical Intelligence used this task to test training-time real-time chunking.
    Photo: Kevin Black et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Knowledge Insulation

  28. A close view of robot grippers at an espresso station, one holding a portafilter filled with ground coffee next to a tamping stand.
    A π0.6 policy making espresso on Physical Intelligence's bimanual rig. The pi line runs one vision-language backbone with a flow-matching action expert across tasks like this one.
    Photo: Kevin Black et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Pi Line

  29. A strip of frames showing two robot arms taking a shirt from a basket, spreading it on a table and folding it.
    Laundry folding is the kind of high-frequency task FAST was built for: with naive per-step tokens, policies made no progress on T-shirt folding at 50 Hz. Frames of the task from the FAST paper.
    Paper figure: Karl Pertsch et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Pi Line

  30. A robot gripper holds a freshly struck match above a pie with a candle in it, while a second gripper holds the open matchbox.
    Real-time chunking lets a robot strike a match and light a candle despite inference delays above 300 ms, more than 30% of the model's prediction horizon. A frame from the RTC paper's candle task.
    Photo: Kevin Black et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Real-Time Execution

  31. An NVIDIA Jetson AGX Orin developer kit, a compact black and silver module with a heat sink, beside its retail box on a white background.
    On-robot compute such as the NVIDIA Jetson AGX Orin developer kit sets the hardware wall: a policy that fits a desktop GPU's budget can miss the control tick here.
    Photo: Auledas / Wikimedia Commons. Licence: CC BY 4.0.

    Appears on: Hardware Taxonomy, Real-Time Execution

  32. Four photographs of robot arms whipping a block out of a Jenga tower, flipping an object in a pan, assembling a timing belt and fitting a car dashboard.
    Reinforcement learning on the real robot reaches precise, dynamic skills that demonstrations alone struggle with. Four of the HIL-SERL tasks: Jenga whipping, object flipping in a pan, timing-belt assembly and dashboard assembly.
    Paper figure: Jianlan Luo et al. / arXiv. Licence: CC BY 4.0.

    Appears on: RL Fine-Tuning of Policies

  33. Photographs of a Franka arm picking up a blue cube and plugging in a yellow peg, and of bimanual arms inserting a GPU card into a motherboard slot in four stages.
    PLD trains a residual RL policy on top of a frozen VLA and distills the improved behavior back into it. On bimanual YAM arms the result ran GPU insertion and unplugging for a continuous hour without human resets.
    Paper figure: Wenli Xiao et al. / arXiv. Licence: CC BY 4.0.

    Appears on: RL Fine-Tuning of Policies

  34. Eight labelled photographs of a Franka arm picking a banana, a spoon and bread, opening a drawer and a toaster, putting bread in a toaster, inserting a chair wheel and hanging a Chinese knot.
    ConRFT fine-tunes a VLA with real-world reinforcement learning across eight contact-rich tasks, from picking a banana to inserting a chair wheel and hanging a Chinese knot.
    Paper figure: Yuhui Chen et al. / arXiv. Licence: CC BY 4.0.

    Appears on: RL Fine-Tuning of Policies

  35. A black Unimate PUMA 500 industrial robot arm from 1983 on a museum plinth, beside its Unimation control cabinet and a computer terminal.
    Every learned policy still runs through coordinate frames like those of this Unimate PUMA 500 (1983), an industrial arm whose kinematics became the textbook example. Exhibit at the Deutsches Museum, Munich.
    Photo: Theoprakt / Wikimedia Commons. Licence: CC BY-SA 3.0.

    Appears on: Robot Learning Curriculum for ML Engineers

  36. A grid of robot camera frames grouped by scene type, showing the same Franka arm working in kitchens, bedrooms, bathrooms, offices, laundry rooms and other real buildings.
    Frames from DROID grouped into its 9 scene types. The same Franka setup recorded every trajectory, in kitchens, bedrooms, offices and other real buildings.
    Paper figure: Alexander Khazatsky et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Data Bottleneck, Robot Learning Curriculum for ML Engineers

  37. A grid of rendered simulated kitchens in different layouts and styles, each with a robot standing among cabinets, counters and appliances.
    Simulation lets a team test policies across many environments before touching hardware. Rendered RoboCasa kitchens, each holding a simulated robot among varied layouts and styles.
    Paper figure: Soroush Nasiriany et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Robot Learning Curriculum for ML Engineers, Generative Simulation

  38. Two rows of frames of a simulated four-legged robot, labelled True and Model: two context frames, then open-loop predictions from step 10 to step 50 in which the model row closely tracks the true row.
    A world model predicts how an environment evolves under actions. From 5 context frames and the actions alone, DreamerV3 predicts 45 frames of a simulated quadruped; the model row tracks the true row.
    Paper figure: Danijar Hafner et al. / arXiv. Licence: CC BY 4.0.

    Appears on: What Is a World Model?

  39. A photorealistic rendered view of a garden: a round wooden table with a vase on it, a paved patio, shrubs and a brick wall.
    A 3D Gaussian Splatting reconstruction is a learned renderer: it predicts a real scene from a new camera pose by rasterising the fitted Gaussians. A garden scene rendered by the original method.
    Paper figure: Bernhard Kerbl et al. / arXiv. Licence: CC BY 4.0.

    Appears on: What Is a World Model?, World Models vs Physics Simulators

  40. A grid of rendered simulation tasks: red robot arms reaching, pushing and opening objects on tables, articulated stick figures on checkered floors, and robotic hands manipulating objects.
    Latent dynamics models learn a compact state and plan or train inside it. TD-MPC2 plans through such a model with no image decoder; these are initial states from its 104 continuous-control tasks.
    Paper figure: Nicklas Hansen et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Latent-Dynamics World Models

  41. A UR5 robot arm with a black gripper above two bins on a table, with three coloured balls in the bins and a colourful picture behind them.
    DayDreamer's UR5 learned multi-object visual pick and place from third-person camera images and sparse rewards within a few hours of autonomous operation, training a Dreamer world model on the real robot without a simulator.
    Photo: Philipp Wu et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Latent-Dynamics World Models

  42. A black and yellow Unitree A1 quadruped robot lying on its side on a grey floor, legs bent, with a cable trailing behind it.
    With a Dreamer-style world model, a Unitree A1 that started on its back learned to roll over, stand up and walk in about one hour of real-world training. A frame from the DayDreamer experiment.
    Photo: Philipp Wu et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Model-Based Robot Learning

  43. Two real tabletop robot images with instructions, wipe the table and remove the bowl from the plate, each turned into a row of generated frames set in a spacecraft interior and a disaster site, with a generalist robot policy acting in the loop.
    A neural simulator can stage scenes no lab has. RoboWorld turns real robot images into spacecraft interiors and disaster sites and runs closed-loop policy evaluation inside the generated rollouts.
    Paper figure: Byeongguk Jeon et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Model-Based Robot Learning

  44. Three pairs of frame strips for the instructions grasp the blue block and rotate it right, put the broccoli onto the plate, and open the drawer; green-framed ground-truth frames sit above blue-framed predicted frames.
    A video world model predicts future frames of a robot task from the current view and a language instruction. GR-1's predicted frames (blue) beneath the ground truth (green) for three instructions.
    Paper figure: Hongtao Wu et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Generative Video World Models

  45. Four groups of tabletop photographs of a robot arm among household objects, labelled Basic Settings, Unseen Backgrounds, Unseen Environments and Unseen Manipulation.
    GR-2 pre-trains on web-scale video, then fine-tunes on robot data to generate video and predict actions, with 97.7% average success across more than 100 tasks. Its test settings run from basic scenes to unseen backgrounds and environments.
    Paper figure: Chi-Lam Cheang et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Generative Video World Models

  46. Three blocks of generated video frames of a robot gripper and objects, each with its instruction and a short evaluator note; after contact, the manipulated object breaks apart or turns into a yellow smear.
    Video world models still fail at contact: after the robot touches it, the manipulated object disintegrates or morphs into artifacts. RoboWorld's VLM evaluator catches such rollouts and gives them a dedicated score.
    Paper figure: Byeongguk Jeon et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Generative Video World Models

  47. A grid of five rows of camera frames of a white Franka robot arm on a cluttered lab bench, in columns labelled Start Frame, Subgoal 1, Subgoal 2 and Goal Frame, as the arm picks up a cup and places it.
    V-JEPA 2-AC plans in embedding space: on Franka arms in labs absent from its training data, it reaches image goals by comparing predicted representations. Pick and place adds two subgoal images between the start and goal frames.
    Paper figure: Mido Assran et al. / arXiv. Licence: CC BY 4.0.

    Appears on: JEPA and the Non-Generative Counterargument

  48. A flow diagram: internet video and images feed video pretraining of V-JEPA 2, which branches to language alignment, attentive probe training and, with robot data of states and actions, action-conditioned post-training for planning and robot manipulation.
    V-JEPA 2 learns from video first and actions second: masked representation prediction on over 1 million hours of internet video and images, then an action-conditioned predictor trained on under 62 hours of robot video.
    Paper figure: Mido Assran et al. / arXiv. Licence: CC BY 4.0.

    Appears on: JEPA and the Non-Generative Counterargument

  49. Two panels: real robot evaluation, labelled expensive, slow and difficult to reproduce, with photographs of a robot arm at a table; and simulated evaluation, labelled cheap, scalable and fully reproducible, with matching simulated views.
    A cheaper evaluator is useful only if it agrees with reality. SIMPLER builds simulated scenes matched to real ones and checks whether simulated policy performance correlates with real performance; a learned simulator should meet the same test.
    Paper figure: Xuanlin Li et al. / arXiv. Licence: CC BY 4.0.

    Appears on: World Model Evaluation for Robotics

  50. A grid of video frames of a robot arm pushing objects at steps 0, 8, 14, 21 and 29, with rows for ground truth and five rollout methods; red boxes mark misaligned and blurred frames, and the last row is marked aligned.
    Autoregressive rollouts can drift in ways one frame hides. On BAIR robot pushing, the RoboWorld authors mark other forcing schemes misaligned or blurred within 29 steps, while their step forcing stays aligned with the ground truth.
    Paper figure: Byeongguk Jeon et al. / arXiv. Licence: CC BY 4.0.

    Appears on: World Model Evaluation for Robotics

  51. A rendered kitchen table crowded with hundreds of 3D objects: fruit, vegetables, meat, bottles, cans, cups, cookware and packaged food.
    Generative simulation authors content and leaves dynamics to the physics engine. RoboCasa's library holds 2,509 objects in 153 categories, drawn from Objaverse and the text-to-3D service Luma.ai; a small subset is shown here.
    Paper figure: Soroush Nasiriany et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Generative Simulation

  52. Rows of identical simulated white robot arms on grey tables, each with a small green part to assemble, with two insets enlarging a gear and a peg.
    A physics simulator integrates explicit geometry, dynamics and contact rules, and steps many copies at once. Assembly environments in Isaac Lab, each an arm with a part to insert, run in parallel on the GPU.
    Paper figure: NVIDIA et al. / arXiv. Licence: CC BY 4.0.

    Appears on: World Models vs Physics Simulators

  53. A white robot arm with a gripper above a plate of parts, labelled Robot Setup, beside a grid of close-up photographs of green parts inserted into black fixtures, labelled Successful Assemblies.
    Either kind of simulator earns trust only against the same real tasks. The Isaac Lab assembly tasks on a real arm, with its successful assemblies.
    Paper figure: NVIDIA et al. / arXiv. Licence: CC BY 4.0.

    Appears on: World Models vs Physics Simulators

  54. A grey and light blue Universal Robots UR16e arm with six revolute joints, folded over a perforated steel table beside its controller box and a touchscreen teach pendant.
    A Universal Robots UR16e, a six-joint serial arm. Forward kinematics composes one transform per joint, so its six joint angles fix the pose of the tool flange.
    Photo: Auledas / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Kinematics

  55. A white two-armed PR2 robot stands at a round red table in an office, reaching toward small objects, with a whiteboard and computer monitors behind it.
    A Willow Garage PR2 in 2010. The TrajOpt benchmark below planned 198 arm problems and 96 full-body problems for this robot in simulated MoveIt scenes.
    Photo: Steve Jurvetson / Wikimedia Commons. Licence: CC BY 2.0.

    Appears on: Motion Planning

  56. A simulated black Atlas humanoid robot stands at a table in a grey virtual warehouse, holding a red hose toward a green valve wheel on a wall.
    A 2013 Defense Department rendering of a simulated Atlas connecting a hose, a DARPA Robotics Challenge task. TrajOpt planned statically stable walking for this 34-DoF humanoid.
    Photo: U.S. Department of Defense / Wikimedia Commons. Licence: Public domain.

    Appears on: Motion Planning

  57. A cast-iron centrifugal governor: two heavy balls hang from arms on a vertical spindle topped by a coiled spring, linked by levers toward the engine valve gear.
    The centrifugal governor of the mill engine at Queen Street Mill, Burnley. As the engine speeds up the balls swing outward and the linkage closes the steam valve, a mechanical feedback loop.
    Photo: Clem Rutter / Wikimedia Commons. Licence: CC BY 3.0.

    Appears on: Control

  58. The hydraulic Atlas humanoid robot stands in a lab during testing, with its exposed metal frame, hoses and sensor head visible and its forearms ending without hands.
    The hydraulic Atlas during 2013 testing, before its hands were fitted. Kuindersma and colleagues built its controller from optimizations running from footstep placement down to whole-body control.
    Photo: DARPA / Wikimedia Commons. Licence: Public domain.

    Appears on: Control

  59. A Rethink Robotics Baxter, a two-armed red and black robot with a screen showing cartoon eyes, stands at a trade-show booth.
    Rethink Robotics' Baxter at Innorobo 2015. Its joints drive through series elastic actuators, so a measured spring deflection gives the arm the force channel a stiff geared arm lacks.
    Photo: Scailyna / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Control

  60. An Apollo inertial measurement unit on display with its round outer case open, showing the gimballed stable platform with its gyroscopes and accelerometers inside.
    An Apollo inertial measurement unit at Draper's 2019 Hack the Moon exhibit. Its gyroscopes and accelerometers integrate motion and drift over time, so an estimator must correct them with outside measurements.
    Photo: ArnoldReinhold / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: State Estimation

  61. The Apollo Guidance Computer, a flat gold-coloured metal case, beside its display and keyboard unit with a numeric keypad and status lights, on a blue background.
    The Apollo Guidance Computer and its display and keyboard. Apollo used Potter's square-root form of the Kalman filter, with ground radar as primary navigation and the onboard system as backup.
    Photo: NASA / Wikimedia Commons. Licence: Public domain.

    Appears on: State Estimation

  62. A Care-O-bot service robot with a white and dark grey body extends an orange jointed arm and a multi-fingered hand to grasp a white cylinder on a folding table.
    A Care-O-bot service robot grasping a cylinder in 2010. A grasp planner chooses where its fingertips touch so the hold resists pushes and pulls from every direction.
    Photo: Jiuguang Wang / Wikimedia Commons. Licence: CC BY-SA 2.0.

    Appears on: Grasp Planning

  63. A black five-fingered Schunk SVH robot hand silhouetted against a window, fingers spread, with the joints and links of each finger visible.
    The Schunk SVH, a servo-electric five-finger hand. Every fingertip or phalanx that touches an object is one contact, and the contact model fixes which forces it can transmit.
    Photo: NearEMPTiness / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Grasp Planning

  64. An Intel RealSense D435 depth camera, a slim grey bar with two infrared lenses, a projector window and a colour camera, mounted on a small tripod.
    An Intel RealSense D435, an active stereo camera common on manipulation rigs. Its projector adds texture so matching works on blank surfaces, and its depth error grows with the square of distance.
    Photo: Marc Auledas / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Perception for Manipulation

  65. A black Kinect for Xbox One sensor bar, angled on a white background, with its camera window at one end and a ventilation grille along its top.
    The Kinect for Xbox One measured depth by time of flight, the principle its successor Azure Kinect kept. Multi-path light and weak infrared returns invalidate pixels on both.
    Photo: Evan-Amos / Wikimedia Commons. Licence: Public domain.

    Appears on: Perception for Manipulation

  66. A white and black Leica laser tracker on a tripod aims a red laser line at the inside of a large rust-coloured steel ring in a factory hall.
    A Leica ATS600 laser tracker measuring a large steel part. Kinematic calibration relies on external instruments like this for its measurement step, recording where the tool really is.
    Photo: FanchonH / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Robot Calibration

  67. Close view of a red Baxter robot arm's wrist, with a small camera lens above a two-finger parallel gripper, in an office with computers behind it.
    Baxter's wrist carries a camera beside its gripper. Hand-eye calibration solves AX = XB for the fixed transform between that camera and the gripper.
    Photo: Gustavo A. Casañ / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Robot Calibration

  68. A white two-armed PR2 robot with raised grippers stands at a Maker Faire booth under banners for PR2, TurtleBot and ROS.
    A Willow Garage PR2 under the ROS banner at Maker Faire 2011. Willow Garage developed ROS alongside this robot, and ROS 2 is its rebuild for production use.
    Photo: Timothy Vollmer / Wikimedia Commons. Licence: CC BY 2.0.

    Appears on: ROS 2 for Machine Learning Engineers

  69. A small black and yellow four-wheeled Clearpath Jackal ground robot with a cylindrical lidar and cameras on its top deck, parked on gravel.
    A Clearpath Jackal at the University of Texas robotics center in 2020. Wheeled research bases like it run Nav2, whose planner, controller and recovery servers are ordinary ROS 2 nodes.
    Photo: U.S. Army / Luke J. Allen / Wikimedia Commons. Licence: Public domain.

    Appears on: ROS 2 for Machine Learning Engineers

  70. An oblique overhead view of a city intersection drawn entirely in coloured lidar points, with building facades, parked cars and road markings in purple, pink and orange on black.
    About 23 million lidar points of a San Francisco intersection, captured in 18 seconds by an Ouster OS1 on a moving car and registered by SLAM. A point cloud is the first rung of the ladder below.
    Photo: Daniel L. Lu / Wikimedia Commons. Licence: CC BY 4.0.

    Appears on: Scene Representation and Mapping

  71. A black Kinect for Xbox 360 sensor bar on its pivoting stand, lying on a green cutting mat with its cables coiled beside it.
    A first-generation Kinect for Xbox 360. KinectFusion fused this sensor's 640×480 depth frames at 30 Hz into a truncated signed-distance volume on one GPU.
    Photo: Antonio Roberts / Wikimedia Commons. Licence: CC BY-SA 2.0.

    Appears on: Scene Representation and Mapping

  72. A Velodyne HDL-64E lidar spinning on a car roof, its cylindrical housing blurred by rotation, above a yellow laser warning label.
    A Velodyne HDL-64E on Junior 2, Stanford's autonomous Volkswagen, in 2009. Its 64 spinning beams return over a million points a second, the raw input of a lidar SLAM front end.
    Photo: Steve Jurvetson / Wikimedia Commons. Licence: CC BY 2.0.

    Appears on: Scene Representation and Mapping

  73. The electric Atlas humanoid by Boston Dynamics, grey and white with a round head and slim jointed limbs, stands on display in a showroom.
    The electric Atlas on display at Hyundai Motorstudio Goyang in July 2026. ZEST trains motion-imitation policies on motion capture, video and animation and deploys them on this robot zero-shot.
    Photo: Damian B Oh / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Humanoid Whole-Body Control

  74. Performers in black motion capture suits dotted with reflective markers stand in a studio capture volume ringed by cameras, seen mostly from behind.
    Optical motion capture records body movement from reflective suit markers seen by a ring of cameras. AMASS pooled 15 such datasets, over 40 hours and 11,000 motions, into one body model that humanoid tracking policies learn to imitate.
    Photo: Applearts123 / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Humanoid Whole-Body Control

  75. Unitree robots on a dark exhibition stand: a silver humanoid hangs from a lifting hoist in front, a G1 humanoid with a glowing blue face ring stands behind it, and a quadruped sits on the right.
    Unitree robots at Japan Mobility Show 2025, with a G1 behind the hoisted humanoid. GR00T N1.7 drives the G1 through GEAR-SONIC, which decodes its latent action tokens into leg, arm and hand commands at 50 Hz.
    Photo: RuinDig/Yuki Uchida / Wikimedia Commons. Licence: CC BY 4.0.

    Appears on: Humanoid Whole-Body Control

  76. A yellow and black Boston Dynamics Spot quadruped walks along a gravel path through a forest.
    Spot, the Boston Dynamics quadruped, outdoors in 2023. Boston Dynamics ships its learned locomotion policy to Spot by software update and reports fewer falls on slick and irregular ground.
    Photo: Jonte / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Legged Locomotion Lineage

  77. A card of twelve sequential silhouette photographs of a horse and rider at a gallop, taken by Eadweard Muybridge in 1878, with printed text beneath the frames.
    Eadweard Muybridge's 1878 sequence of a galloping horse caught all four hooves off the ground. A gait is such a footfall timing pattern; Walk These Ways commands it to a robot as phase offsets between foot pairs.
    Photo: Eadweard Muybridge / Wikimedia Commons. Licence: Public domain.

    Appears on: Legged Locomotion Lineage

  78. Cassie, a bipedal robot with backward-bending legs and no upper body, walks along a forest path while a researcher follows with a handheld controller.
    Cassie, the bipedal robot built at Oregon State University, walking in a research forest in 2017. Learned controllers later ran Cassie through a 400-meter dash and composed periodic rewards into standing, walking, hopping, running and skipping.
    Photo: Oregon State University / Wikimedia Commons. Licence: CC BY-SA 2.0.

    Appears on: Legged Locomotion Lineage

  79. An ANYmal X quadruped robot climbs down a steel staircase on an offshore platform, surrounded by railings and industrial piping.
    ANYmal X descending stairs on an offshore rig during an inspection round in 2022. The ANYmal line is where simulation-trained locomotion policies first transferred to hardware, in 2019.
    Photo: Cvoellmy / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Why RL Won Locomotion but Not Manipulation

  80. The Shadow Dexterous Hand, a humanlike robot hand with visible joints and tendons, holds a light bulb between its fingertips.
    The Shadow Dexterous Hand holding a light bulb in 2007. Each fingertip is a contact whose geometry and friction a simulator must get right; a modified Shadow hand ran the Rubik's cube policy trained with automatic domain randomization.
    Photo: Richard Greenhill and Hugo Elias, Shadow Robot Company / Wikimedia Commons. Licence: CC BY-SA 3.0.

    Appears on: Why RL Won Locomotion but Not Manipulation

  81. A rendered Shadow Hand holding a cube in the foreground, with rows of identical hand and object environments receding behind it across a grey checkered floor.
    Isaac Gym's Shadow Hand task, rendered with many copies of the same hand and cube stepped together on one GPU. Its authors report 2 to 3 orders of magnitude improvement over a CPU simulator feeding a GPU network.
    Paper figure: Viktor Makoviychuk et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Massively Parallel Sim RL

  82. Rendered robots from MuJoCo Playground arranged together: a Berkeley Humanoid, Unitree Go1 and G1, a LEAP hand and a Franka arm.
    MuJoCo Playground environments that transferred to real hardware: the Berkeley Humanoid, Unitree Go1 and G1, the LEAP hand and a Franka arm. The framework runs MuJoCo on the GPU through MJX, with sim-to-real transfer as an express goal.
    Paper figure: Kevin Zakka et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Massively Parallel Sim RL

  83. The MIT Mini Cheetah, a small silver quadruped on a safety tether, spins in place on foam floor mats, its legs blurred by motion.
    The MIT Mini Cheetah spinning at 5.7 rad/s under a policy trained in simulation, 90% of the 6.28 rad/s yaw rate a model-based controller reached on the robot. The same network sprints at 3.9 m/s.
    Photo: Gabriel B. Margolis et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Reward Design and the MPC Debate

  84. A rendered grid of terrain tiles with stairs, slopes and rough ground, covered with many ANYmal quadrupeds training at once against a black background.
    ANYmal robots training in parallel on Isaac Gym stairs and slopes, laid out as tiles of rising difficulty. The curriculum promotes robots that cross a tile and demotes those that cover under half the commanded distance.
    Paper figure: Viktor Makoviychuk et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Reward Design and the MPC Debate

  85. Overlaid frames show a Unitree A1 quadruped on a lab floor being pushed with a long tube and keeping or regaining its footing.
    DayDreamer's Unitree A1 learned to roll over, stand and walk in one hour on the real robot with no simulator. Within ten more minutes of pushes it learned to withstand them or roll back onto its feet.
    Photo: Philipp Wu et al. / arXiv. Licence: CC BY 4.0.

    Appears on: RL for Robotics

  86. Rendered Allegro robot hands, each turning a cube, repeated across a large grid of parallel simulation environments on a grey checkered floor.
    Isaac Gym's Allegro hand task, one hand and cube per environment, all stepped together on a GPU. When steps cost this little, training stability becomes the scarce resource, which is why GPU-simulation work defaulted to PPO.
    Paper figure: Viktor Makoviychuk et al. / arXiv. Licence: CC BY 4.0.

    Appears on: RL for Robotics

  87. A rendered TriFinger robot: three robotic fingers mounted above a round, walled arena push a cube toward a target pose.
    TriFinger, three robot fingers over a round arena, rendered in Isaac Gym. The real benchmark pairs reproducible hardware with simulation for debugging, so offline-RL policies trained on its fixed datasets meet the same robot every time.
    Paper figure: Viktor Makoviychuk et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Offline Reinforcement Learning for Robotics

  88. Three robot scenes in four columns labelled Real, Raw Sim, Sim plus Green Screen, and Sim plus Green Screen plus Texture Matching; the simulated views grow closer to the real camera images from left to right.
    SIMPLER matches its simulated scenes to real camera views, from raw simulation to a green-screened real background with matched textures. The authors checked simulated rankings against paired real evaluations before trusting them.
    Paper figure: Xuanlin Li et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Offline Reinforcement Learning for Robotics

  89. A red and black ANYmal quadruped climbs a short flight of carpeted steps inside an office building, with wooden blocks on the floor nearby.
    A real ANYmal climbing indoor steps with a rough-terrain policy trained in Isaac Gym. The Isaac Gym paper shows the same policy on simulated terrain beside this hardware test.
    Photo: Viktor Makoviychuk et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Sim-to-Real Transfer

  90. A real Franka Emika Panda arm with a Robotiq gripper pushes a foam yoga block across a table toward a marked goal region.
    A Franka Panda pushes a yoga block to a goal with a MuJoCo Playground policy trained under domain randomization and transferred zero-shot. Early policies moved fast enough to trip the arm's force limit, so the authors added torque penalties.
    Photo: Kevin Zakka et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Sim-to-Real Transfer

  91. A scan, label, apply and manifest machine straddling a parcel conveyor inside Amazon's Swindon fulfilment centre, with label printers mounted above the belt inside a yellow guard frame.
    A parcel labelling line at Amazon's Swindon fulfilment centre in 2025. Fixed automation repeats one narrow task and has to finish it nearly every time, the product end of the gap.
    Photo: Auledas / Wikimedia Commons. Licence: CC BY 4.0.

    Appears on: The Reliability Gap

  92. Two silver and black Tesla Optimus humanoid robots standing on round black pedestals in a white-walled showroom.
    Two Tesla Optimus humanoids on display in Stockholm in 2024. Tesla's 2026 updates list both of its Optimus production lines as still under construction.
    Photo: Ulf Klingström / Wikimedia Commons. Licence: Public domain.

    Appears on: The Reliability Gap

  93. Honda's white ASIMO humanoid robot standing on a polished floor beneath the large suspended globe display at Miraikan, Tokyo.
    Honda's ASIMO performing at Miraikan, Tokyo, in 2016. The bear case holds that capital arrives on demonstrations like this one and deployment evidence arrives late.
    Photo: Sylvain Kalache / Wikimedia Commons. Licence: CC BY 2.0.

    Appears on: The Bear Case

  94. A round grey and black iRobot Roomba 870 robot vacuum cleaner on a light wooden floor.
    An iRobot Roomba 870 in 2014. Rodney Brooks co-founded iRobot, which makes the Roomba, so his forecasts come from someone who has shipped robots.
    Photo: Kārlis Dambrāns / Wikimedia Commons. Licence: CC BY 2.0.

    Appears on: The Bear Case

  95. On a dark, spotlit stage, a Unitree G1 humanoid, a black four-legged FOX robot and a small Unitree Go2 robot dog stand in a row on a white floor.
    A Unitree G1 humanoid on stage at the Japan Mobility Show 2025. Unitree is one of five backers of the humanoid thesis, one of six rival bets on how robot intelligence gets built.
    Photo: RuinDig/Yuki Uchida / Wikimedia Commons. Licence: CC BY 4.0.

    Appears on: Competing Theses

  96. An orange Amazon warehouse drive unit, a low square wheeled robot with a black circular lifting plate on top, standing on display.
    An Amazon warehouse drive unit, photographed in 2023. Amazon has deployed its millionth robot, and the task-specific side bets that purpose-built machines like this one win on cost per task.
    Photo: Geni / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Competing Theses

  97. A four-fingered Allegro robot hand on a white background, every finger segment and fingertip covered in soft black tactile sensor pads.
    XELA Robotics' uSkin tactile sensors on every finger segment of an Allegro hand, 2021. They measure the contact that cameras lose once the fingers close.
    Photo: XELA Robotics / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Dexterity

  98. A white bimanual robot with two dexterous hands at a wooden table, shown from the front and the side, with callouts for its head camera and wrist cameras and, at left, the Manus gloves, Vive trackers and laptop controller used to teleoperate it.
    The T-Rex rig: a bimanual robot with two 22-DoF Sharpa Wave hands, a head camera and two wrist cameras, and the gloves and trackers that teleoperate it.
    Paper figure: Dantong Niu et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Dexterity

  99. Six photographs of one robotic hand and a 2×2×2 cube, labelled Pick up, then Start, Turn, Complete and Reset for one 90-degree layer turn, then Solved.
    FINGR picks up a 2×2×2 cube, turns one layer at a time inside a single real hand and resets its fingers between turns until the cube is solved.
    Paper figure: Yutong Liang et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Dexterity

  100. A grid of robot camera frames in a home kitchen: rows labelled DROID (ours), OXE and No Co-Train, columns labelled Remove Lid, Pick up Lentils, Pour Lentils and Turn on Stove, each frame outlined green for success or red for failure.
    One cooking task in a home kitchen with unseen distractors, under three training mixes. Only the policy co-trained on DROID's in-the-wild data finishes all four steps.
    Paper figure: Alexander Khazatsky et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Generalization

  101. Two humanoid robots side by side: a Galaxea R1 Pro labelled 22-dof Sharpa Hand and a Unitree G1 labelled 7-dof Tri-finger Hand, with lines to the OAK-D-Wide head camera and OAK-1-Wide wrist cameras each carries.
    EgoScale's two robots: a Galaxea R1 Pro with 22-DoF Sharpa hands and a Unitree G1 with 7-DoF tri-finger hands. On the G1, human-video pretraining added over 30 points of absolute success on both evaluated tasks.
    Paper figure: Ruijie Zheng et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Generalization

  102. A robotic welding cell on a factory floor, seen from above, enclosed by red weld curtains and yellow and black guard posts.
    A robotic weld cell behind machine guarding and weld curtains. Fencing people out is the oldest safeguard; robots that work without a fence use one of four collaborative modes.
    Photo: WireCrafters / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Safety and Assurance

  103. A red mushroom-head Siemens emergency-stop button on a round yellow collar, set into a dark machine-control panel beside a keypad.
    A Siemens emergency-stop actuator on a machine control panel. It acts only after a person has seen the hazard and pressed it.
    Photo: Stahlkocher / Wikimedia Commons. Licence: CC BY-SA 3.0.

    Appears on: Safety and Assurance

  104. Photograph of a slender robot of white frames linked by pairs of flexible racks, curving across a black background, with labels for Fixed Base, PLA Racks, Embedded Motor and Lighthouse Deck.
    Cai and colleagues' hyper-redundant robot: five segments that extend and bend as their motors drive paired flexible racks. Weighted barrier functions kept it clear of obstacles while correcting its tracking.
    Paper figure: Zijian Cai et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Safety and Assurance

  105. Three-column diagram: a person asks a robot arm to put the bowl in the microwave; the planner scores four next steps, conformal prediction keeps two above its threshold, and the robot asks which one, plastic or metal; photos at right show the test environments.
    Told to put the bowl in the microwave with a plastic and a metal bowl in view, KnowNo keeps two plans above its conformal threshold and asks which bowl is meant.
    Paper figure: Allen Z. Ren et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Safety and Assurance

  106. A white Waymo Jaguar I-Pace with a roof-mounted sensor dome and sensor pods on its fenders driving through a San Francisco intersection, the downtown skyline behind it and nobody in the driver’s seat.
    A Waymo Jaguar I-Pace driving itself through San Francisco in 2023. Waymo's 2025 study of 56.7 million rider-only miles found statistically significant crash reductions against human benchmarks in every outcome it could test.
    Photo: Dllu / Wikimedia Commons. Licence: CC BY 4.0.

    Appears on: Autonomous Vehicles

  107. Close-up of the roof of a white Waymo Jaguar I-Pace: a domed lidar housing on a raised black mount, ringed by smaller camera and radar windows.
    The roof sensor package of a Waymo I-Pace in San Francisco, 2025. Waymo fuses lidar, camera and radar, paying for active range sensing that works in the dark and measures geometry directly.
    Photo: 9yz / Wikimedia Commons. Licence: CC BY 4.0.

    Appears on: Autonomous Vehicles

  108. Front view of Carnegie Mellon's Boss, a black Chevrolet Tahoe numbered 19, with a rack of lidar and camera sensors on its roof and more sensors on its bumper, on display in a darkened hall.
    Carnegie Mellon's Boss, winner of the 2007 DARPA Urban Challenge, on display in 2008. Route planning, behaviour decisions and local motion planning, the classical toolbox of its era, remain live production components.
    Photo: Steve Jurvetson / Wikimedia Commons. Licence: CC BY 2.0.

    Appears on: Autonomous Vehicles

  109. A small racing quadrotor on a wooden table: a bare carbon frame with four orange propellers, a green flight-controller board, a battery strapped on top and a forward-facing camera.
    A Minion 220 racing quadrotor with its forward camera, 2016. Machines of this class exceed 100 km/h, and in 2023 a reinforcement-learning system beat three human world champions at racing them.
    Photo: Commanderbryce / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Drones and Aerial Robotics

  110. A grey DJI Mavic Pro quadcopter hovering against a cloudy sky, propellers blurred, its camera hanging below the front of the body.
    A DJI Mavic Pro in flight, 2017. On an autonomous drone, visual-inertial odometry fuses camera features with inertial measurements into the pose and velocity the controller needs, with no GPS or motion capture.
    Photo: Sinky / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Drones and Aerial Robotics

  111. NASA's Perseverance rover on reddish Martian ground, its mast camera head raised, beside a flat rock with two small drill holes, under a hazy orange sky.
    Perseverance beside the rock Rochette on sol 198, September 2021, with two drill holes where its arm took cores. In its first Mars year, its AutoNav system evaluated 88 percent of the 17.7 km the rover drove.
    Photo: NASA / Wikimedia Commons. Licence: Public domain.

    Appears on: Space Robotics

  112. The Ingenuity helicopter standing on four thin legs on rust-coloured Martian soil, its two stacked rotors and the solar panel on top casting a long shadow.
    Ingenuity on Mars, photographed by Perseverance on 5 April 2021, two weeks before its first flight. Qualified for five flights in thirty days, it flew 72 times over almost three years.
    Photo: NASA/JPL-Caltech/ASU/MSSS / Wikimedia Commons. Licence: Public domain.

    Appears on: Space Robotics

  113. Technicians in white cleanroom suits lower a gold-coloured box, the MOXIE oxygen experiment, into the open chassis of the Perseverance rover in a clean room.
    Mars 2020 team members install MOXIE in Perseverance's chassis. On Mars it made 122 grams of oxygen from atmospheric carbon dioxide across sixteen runs, about twice NASA's original production goals.
    Photo: NASA / Wikimedia Commons. Licence: Public domain.

    Appears on: Space Robotics

  114. The two-armed Dextre robot hanging from the end of the long white Canadarm2 outside the International Space Station, with solar arrays and the black of space behind.
    Dextre on the end of Canadarm2 in January 2017, during weeks spent swapping the station's batteries. Ground crews plan and fly every such task by teleoperation.
    Photo: NASA / Wikimedia Commons. Licence: Public domain.

    Appears on: Space Robotics

  115. A white da Vinci Xi patient-side cart in a hospital room, four jointed instrument arms hanging from an overhead boom on a wheeled base.
    A da Vinci Xi patient-side cart with its four instrument arms, 2018. Every motion its arms make is a surgeon’s command.
    Photo: Alvarogarciamd / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Surgical Robotics

  116. The front of a da Vinci Xi surgeon console: two eyepieces of the stereo viewer above a pair of hand controllers, with an armrest and touch panel below.
    The stereo viewer and hand controllers of a da Vinci Xi surgeon console. Motion scaling and tremor filtering sit between these grips and the instruments, and the surgeon remains planner and controller: Level 0 autonomy.
    Photo: A.BourgeoisP / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Surgical Robotics

  117. Eight round fisheye views from a gripper's camera, each a different office or meeting-room table with a mug or bottle in front of the gripper's purple and white jaws.
    The eight test environments in Lin and colleagues' data-scaling study, seen through the gripper's fisheye camera. None was in the training data, and every task was tested in all eight.
    Paper figure: Fanqi Lin et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Data Bottleneck

  118. Seven labelled camera frames from BridgeData V2 demonstrations in a toy kitchen sink, such as Put Carrot on Plate and Put Eggplant into Pot, each with a WidowX robot's black gripper above the toy objects.
    Seven BridgeData V2 tasks in one toy-sink scene, such as putting a carrot on a plate, each recorded on a WidowX 250.
    Paper figure: Moo Jin Kim et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Major Datasets

  119. The LIBERO overview: rendered scenes of a simulated robot arm at kitchen tables in four suites labelled LIBERO-Spatial, LIBERO-Object, LIBERO-Goal and LIBERO-100 around a procedural-generation step, above five research questions such as distribution shift and task ordering.
    LIBERO's simulated task suites. LIBERO-Spatial changes only the layout, LIBERO-Object only the objects and LIBERO-Goal only the goal, and LIBERO-100 holds 100 more varied tasks.
    Paper figure: Bo Liu et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Evaluation Crisis

  120. An orange robot arm with grey joints holds a cast-metal part in a three-fingered robotic hand, against a plain grey backdrop.
    A seven-DoF arm and a three-fingered hand on NIST's Dexterous Manipulation Testbed in 2013, used to measure manipulation for manufacturing tasks.
    Photo: National Institute of Standards and Technology / Wikimedia Commons. Licence: Public domain.

    Appears on: Hardware Taxonomy

  121. Orange KUKA industrial robot arms with welding guns surround a bare metal car body shell on a body-shop line at the BMW plant in Leipzig, with more robots in the hall behind.
    KUKA robots spot-welding BMW 3 Series bodies at the Leipzig plant in 2005. Fixed cells repeating one motion like this make up most of the installed base.
    Photo: BMW Werk Leipzig / Wikimedia Commons. Licence: CC BY-SA 2.0.

    Appears on: Industrial Deployment

  122. Boxy white robots with blue tops run on the rails of an aluminium storage grid that fills an Ocado warehouse hall; one robot on the left is blurred by motion.
    Bots on the storage grid of an Ocado warehouse in 2017. They run on top of the grid and fetch bins from the stacks inside it.
    Photo: Techwords / Wikimedia Commons. Licence: CC BY-SA 4.0.

    Appears on: Industrial Deployment

  123. Orange KUKA robot arms inside a wire-mesh safety fence lift yellow crates of bread from a roller conveyor and stack them in tall piles, in a bakery hall.
    KUKA robots palletising bread crates at a German bakery in 2005. The grippers, conveyors and safety fence around them are part of the cell's cost.
    Photo: KUKA Roboter GmbH / Wikimedia Commons. Licence: Public domain.

    Appears on: Industrial Deployment

  124. The OpenVLA overview figure: a montage labelled 970k Robot Episodes feeds a vision-language-action model built on Llama 2 7B, above photos of three robots it controls, a WidowX at a toy sink, a mobile robot by an apple and a banana, and a Franka with a dustpan.
    OpenVLA's overview traces one path through the stack: 970k robot episodes, a 7B vision-language-action model and closed-loop control of several robots.
    Paper figure: Moo Jin Kim et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Robot Learning Stack

  125. A grid of DROID camera frames from different scenes, each with a purple mask of the robot arm rendered over the real arm in the image.
    DROID frames with the robot's mesh rendered from the estimated camera-to-base transform. Where the mask covers the real arm, the calibration holds; a moved camera shows as a mask that misses.
    Paper figure: Alexander Khazatsky et al. / arXiv. Licence: CC BY 4.0.

    Appears on: The Robot Learning Stack

  126. Seen from behind, a researcher holds the handle of a master-slave manipulator below a thick leaded-glass hot-cell window lit yellow from inside.
    A researcher at Idaho National Laboratory in 2006 works radioactive material through a leaded-glass window, a handle outside driving tongs inside. ALOHA's leader arm uses the same principle.
    Photo: Idaho National Laboratory / Wikimedia Commons. Licence: CC BY 2.0.

    Appears on: Teleoperation Rigs

  127. Four UMI handheld grippers in a row on a white background, each a white 3D-printed body with purple parallel jaws and a black GoPro camera mounted on top.
    The four UMI handheld grippers used in Lin and colleagues' data-scaling study. A Franka arm with matching purple fingers and a wrist GoPro ran the trained policies.
    Paper figure: Fanqi Lin et al. / arXiv. Licence: CC BY 4.0.

    Appears on: Teleoperation Rigs