Comparison Matrix

The comparison matrix is a table of major robot policies, from RT-1 (2022) to π0.7 (2026), compared on eight architectural axes: action representation, horizon, control frequency, backbone, conditioning, cross-embodiment, hierarchy and weights^{Source 1}1 Brohan et al. 2022, RT-1: Robotics Transformer for Real-World Control at Scale. Anthony Brohan, Noah Brown, Justice Carbajal et al., 2022 ^{Source 11}11 Ai et al. 2026, π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities. Bo Ai, Ali Amin, Raichelle Aniceto et al., 2026. Each cell records what its named source discloses.

Paper figure: Octo Model Team / arXiv. Licence: CC BY 4.0.

Set the weights filter to Downloadable.

Use the matrix as a constraint solver

Remove models whose weights or training data the project cannot use (Gemini Robotics On-Device, for example, is reached through a trusted-tester SDK^{Source 19}19 Parada 2025, Gemini Robotics On-Device brings AI to local robotic devices. Carolina Parada, Google DeepMind blog, 2025), then models whose action representation, latency, observations or adaptation-data needs do not fit the robot, and only then compare benchmark performance.

Paper figure: Pranav Atreya et al. / arXiv. Licence: CC BY 4.0.

Compare the survivors under one trial protocol with the same cameras, action horizon and resets, and keep a behavior-cloning or action-chunking baseline. TRI's Large Behavior Model study used blind, randomized trials^{Source 20}20 TRI LBM Team et al. 2025, A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation. TRI LBM Team, Jose Barreiros, Andrew Beaulieu et al., 2025, and RoboArena ranked seven generalist policies more accurately than centralized evaluation through over 600 double-blind pairwise real-robot episodes^{Source 21}21 Atreya et al. 2025, RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies. Pranav Atreya, Karl Pertsch, Tony Lee et al., 2025. The foundation-model guide explains what prior each family contributes, the action-space guide whether its output fits, and the evaluation crisis why protocols resist a shared leaderboard.

The eight axes

Action representation is the output mechanism: continuous regression, discrete action tokens, or a generative head such as diffusion or flow matchingflow matchingFlow matching is a generative modeling recipe that learns a time-dependent vector field carrying noise to data, by regressing against the conditional flow instead of estimating a score. In pi0, which brought it to robot control, an action expert trained this way denoises continuous action chunks. Integrating the field as an ODE for a handful of steps is fast enough for high-frequency control and keeps diffusion policies' multimodality.. Supervision and runtime generation can differ: π0.5 trains with FAST tokens but generates continuous actions with flow matching^{Source 8}8 Physical Intelligence et al. 2025, π0.5: a Vision-Language-Action Model with Open-World Generalization. Physical Intelligence, Kevin Black, Noah Brown et al., 2025.

Horizon separates the planned chunk length $H$, the predicted horizonpredicted horizonThe predicted horizon is the chunk length H: the number of future actions a chunked policy emits in one forward pass, whether or not it executes them all. It is 16 for Diffusion Policy and 50 actions, one second at 50 Hz, for π0. The committed horizon is how many of them run., from the steps executed before re-planning, the committed horizoncommitted horizonThe committed horizon is the number of predicted actions a system executes before a fresh inference replaces them. Diffusion Policy predicts 16 steps and commits 8; π0.7 predicts 50 and commits 15 to 25; original π0, executing synchronously, committed the whole chunk. Real-Time Chunking freezes the first d actions of each new chunk to the values in flight, so the committed horizon can be shorter than the predicted one without pauses.. Diffusion Policy's simulated CNN Push-T reference uses observation, prediction and execution horizons of 2, 16 and 8^{Source 3}3 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023; ACT's 100-step entry is its k=100 chunk-ablation reference^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023; OFT+ predicts and executes 25 actions at 25 Hz on ALOHA^{Source 6}6 Kim et al. 2025, Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. Moo Jin Kim, Chelsea Finn, Percy Liang, 2025; and π0.7 predicts 50 with five denoising steps and executes 15 or 25^{Source 11}11 Ai et al. 2026, π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities. Bo Ai, Ali Amin, Raichelle Aniceto et al., 2026. Real-time chunking generates the next chunk while the current one executes^{Source 18}18 Black et al. 2025, Real-Time Execution of Action Chunking Flow Policies. Kevin Black, Manuel Y. Galliker, Sergey Levine, 2025.

Control frequency needs a named measurement setting. ACT's ALOHA setup runs at 50 Hz with temporal ensemblingtemporal ensemblingTemporal ensembling is ACT's execution scheme for overlapping action chunks: the policy runs every timestep and averages in-flight predictions for the current step with exponential weights. In the pinned ACT reference implementation, predictions ordered oldest to newest get weight exp(-0.01 * i), so the oldest retained prediction has the largest weight. Handoffs are smoother, at the cost of one inference per step instead of per chunk.^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023, RT-1 commands the Everyday Robots platform at 3 Hz^{Source 1}1 Brohan et al. 2022, RT-1: Robotics Transformer for Real-World Control at Scale. Anthony Brohan, Noah Brown, Justice Carbajal et al., 2022, and RT-2's multi-TPU cloud service runs PaLI-X-55B at 1 to 3 Hz^{Source 4}4 Brohan et al. 2023, RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. Anthony Brohan, Noah Brown, Justice Carbajal et al., 2023. OpenVLA reaches about 6 Hz on one RTX 4090, and its Franka fine-tuning setups use 5 Hz and 15 Hz controllers^{Source 5}5 Kim et al. 2024, OpenVLA: An Open-Source Vision-Language-Action Model. Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti et al., 2024.

Gemini Robotics turns about 250 ms of end-to-end latency into a 50 Hz effective control frequency with multi-action chunks^{Source 22}22 Gemini Robotics Team et al. 2025, Gemini Robotics: Bringing AI into the Physical World. Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie et al., 2025, GR00T N1 samples a 16-action chunk in 63.9 ms on an L40^{Source 23}23 NVIDIA et al. 2025, GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. NVIDIA, Johan Bjorck, Fernando Castañeda et al., 2025, the π0.6 card reports 63 ms per chunk on one H100^{Source 10}10 Physical Intelligence 2025, π0.6 Model Card. Physical Intelligence, 2025, and Generalist says GEN-1.5 produces 100 Hz action trajectories^{Source 24}24 Generalist Team 2026, GEN-1.5: Embodied Foundation Models are One-Shot Learners. Generalist Team, Generalist AI blog, 2026.

Backbone, conditioning and cross-embodiment describe model components, inputs and the scope of robot-body transfer. Mixed-embodiment training, one checkpoint on several bodies and fine-tuning on a new body are different evidence; CrossFormer shows the strongest kind, with the same weights controlling very different robots after training on 900K trajectories from 20 embodiments^{Source 25}25 Doshi et al. 2024, Scaling Cross-Embodied Learning: One Policy for Manipulation, Navigation, Locomotion and Aviation. Ria Doshi, Homer Walke, Oier Mees et al., 2024. Skild says S1 executes a task shown in one video^{Source 26}26 Skild AI Team 2026, Introducing S1: In-Context Learning for Robotics. Skild AI Team, Skild AI blog, 2026, which adds demonstration video as a conditioning input.

Hierarchy distinguishes an internal high-level stage from a separately paired system. π0.5 predicts a semantic subtask before low-level actions^{Source 8}8 Physical Intelligence et al. 2025, π0.5: a Vision-Language-Action Model with Open-World Generalization. Physical Intelligence, Kevin Black, Noah Brown et al., 2025, while in the N1.7 SONIC workflow a separate whole-body controller decodes the VLA's latent action tokens into joint commands^{Source 14}14 NVIDIA 2026, NVIDIA Isaac GR00T. NVIDIA, 2026. A separate high-level VLM gives π0 subtasks^{Source 7}7 Black et al. 2024, π0: A Vision-Language-Action Flow Model for General Robot Control. Kevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024.

Weights records download availability, separately from license. openpi lists π0, π0-FAST and π0.5 checkpoints^{Source 9}9 Physical Intelligence 2026, openpi. Physical Intelligence, 2026; the N1.7 README separates Apache 2.0 code from weights under the NVIDIA Open Model License^{Source 14}14 NVIDIA 2026, NVIDIA Isaac GR00T. NVIDIA, 2026; and SmolVLA releases code, pretrained models and training data^{Source 27}27 Shukor et al. 2025, SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics. Mustafa Shukor, Dana Aubakirova, Francesco Capuano et al., 2025.

Row notes

RT-1's "limited" cross-embodiment label means Kuka IIWA data mixed with Everyday Robots data and evaluated on Everyday Robots^{Source 1}1 Brohan et al. 2022, RT-1: Robotics Transformer for Real-World Control at Scale. Anthony Brohan, Noah Brown, Justice Carbajal et al., 2022. ACT trains a roughly 80M-parameter model from scratch for each task^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023, and RT-2's action-only entry excludes its chain-of-thought variant^{Source 4}4 Brohan et al. 2023, RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. Anthony Brohan, Noah Brown, Justice Carbajal et al., 2023.

Paper figure: Gemini Robotics Team / arXiv. Licence: CC BY 4.0.

OpenVLA's 7B label names the language model, beside a 600M visual encoder^{Source 5}5 Kim et al. 2024, OpenVLA: An Open-Source Vision-Language-Action Model. Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti et al., 2024. OFT's "limited" label means adapting the pretrained model to separate tasks and robot setups^{Source 6}6 Kim et al. 2025, Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. Moo Jin Kim, Chelsea Finn, Percy Liang, 2025, and GR00T N1.6 predicts state-relative action chunks for most embodiments^{Source 28}28 GEAR Team et al. 2025, GR00T N1.6: An Improved Open Foundation Model for Generalist Humanoid Robots. GEAR Team, Allison Azzolini, Johan Bjorck et al., 2025. π*0.6 adds Recap post-training, which more than doubles throughput on some of the hardest tasks^{Source 29}29 Amin et al. 2025, π*0.6: a VLA That Learns From Experience. Ali Amin, Raichelle Aniceto, Ashwin Balakrishna et al., 2025.

Gemini Robotics 1.5's model card discloses continuous numerical action outputs but no action-head family, chunk length, command rate or parameter count; one checkpoint runs across ALOHA, bi-arm Franka and Apollo, and a separate GR-ER 1.5 orchestrator can pair with it^{Source 12}12 Gemini Robotics Team et al. 2025, Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer. Gemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan et al., 2025. DeepMind's July 30, 2026 Gemini Robotics 2 announcement covers Apollo 2 and Franka Duo with one checkpoint but gives no backbone, output parameterization, chunk length, control frequency or parameter count^{Source 13}13 Parada 2026, Gemini Robotics 2 brings whole body intelligence to robots. Carolina Parada, 2026. Skild's row rests on its January 14, 2026 Series C announcement, with no reproducible cross-embodiment evaluation or runtime details^{Source 17}17 Skild AI Team 2026, Announcing Series C. Skild AI Team, 2026.

What the matrix can establish

ACT fixes its CVAE latent to the prior mean, zero, to decode deterministically^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023, while π0 generates continuous action chunks with a flow-matching action expert^{Source 7}7 Black et al. 2024, π0: A Vision-Language-Action Flow Model for General Robot Control. Kevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. Generalist says GEN-1's base model trains on wearable human data without robot data and reaches 99% average success where earlier models reached 64%, with 1 hour of robot data per task^{Source 30}30 Generalist Team 2026, GEN-1: Scaling Embodied Foundation Models to Mastery. Generalist Team, Generalist AI blog, 2026. On March 16, 2026, NVIDIA previewed GR00T N2, based on DreamZero research^{Source 31}31 NVIDIA 2026, NVIDIA and Global Robotics Leaders Take Physical AI to the Real World. NVIDIA, NVIDIA Newsroom, 2026.

Helix 02 and GO-2 disclosure limits

Figure's Helix 02 announcement gives S1 joint targets at 200 Hz and S0 actuator commands at 1 kHz, but no output-head training loss, chunk horizon or weight-release terms^{Source 15}15 Figure AI 2026, Introducing Helix 02: Full-Body Autonomy. Figure AI, 2026. The first Helix post disclosed a 7 to 9 Hz onboard VLM and an 80M-parameter S1 at 200 Hz, trained end-to-end with a standard regression loss^{Source 32}32 Figure AI 2025, Helix: A Vision-Language-Action Model for Generalist Humanoid Control. Figure AI, 2025, and Figure reports that Index human-video pretraining raised Helix 2.5's zero-shot success in 30 unseen homes from 9% to 56%^{Source 33}33 Figure AI 2026, Helix 2.5: Zero-Shot 30-Home Generalization. Figure AI, Figure news, 2026. AgiBot describes lower-frequency System 2 planning and higher-frequency System 1 action following, with no numerical rate, backbone, output encoding, chunk length or weight terms^{Source 16}16 AgiBot 2026, The Unity of Reasoning and Action: AGIBOT Unveils Genie Operator-2 (GO-2) Next-Gen Embodied Foundation Model. AgiBot, 2026; its predecessor GO-1 uses latent action representations trained on AgiBot World's more than 1 million trajectories across 217 tasks^{Source 34}34 AgiBot-World-Contributors et al. 2025, AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems. AgiBot-World-Contributors, Qingwen Bu, Jisong Cai et al., 2025.