Other Generalist Policies

The last two modules each followed one lineage: the VLA family from RT-1 to OpenVLA, then Physical Intelligence's pi line. The rest of the frontier belongs to neither. This module surveys the other generalist policies an ML engineer will actually meet in 2026: Google DeepMind's Gemini Robotics, NVIDIA's GR00T, Figure's Helix, AgiBot's GO series, and Skild. The organizing axis here is access rather than architecture: which ones you can download, and how each one's claims are documented. Keep two questions in mind throughout. Are the weights open? And is the evidence a paper, a repo, a lab blog, or a press release?

The timeline separates reported weight downloads, records marked not downloadable, and availability not disclosed in the inspected source. These are not license classifications. Node labels and filters distinguish the states; marker shape identifies the source format. For Gemini 1.5, Gemini 2, and Skild, read the source-scoped availability note in the selected detail.

Gemini Robotics: the strongest closed line, the thinnest disclosure

The March 2025 Gemini Robotics report introduces a Gemini 2.0-based model family. Gemini Robotics-ER extends spatial and temporal understanding for tasks including pointing, trajectory prediction, and grasp prediction; Gemini Robotics builds on ER to produce robot actions Gemini Robotics Team 2025↓Gemini Robotics: Bringing AI into the Physical WorldGemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie et al., 2025. The 1.5 technical report, first submitted to arXiv on October 2, 2025, describes three advances: natural-language thinking traces interleaved with VLA actions when thinking is enabled; Motion Transfer, a training recipe for learning across robot embodiments; and the separate GR-ER 1.5 embodied-reasoning model used as an orchestrator. The report varies ER's thinking-token budget and finds that the useful amount of thinking depends on the task. The submission date is a report date, not an established product-release date. Gemini Robotics Team 2025↓Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion TransferGemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan et al., 2025 Both releases have technical reports on arXiv, which puts them a tier above most of this module in provenance.

Google DeepMind announced Gemini Robotics 2 on July 30, 2026 as three models: the Gemini Robotics 2 VLA converts vision and language into motor control; Gemini Robotics ER 2 is a separate embodied-reasoning VLM; and Gemini Robotics On-Device 2 is a VLA optimized for local operation. Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026

The announcement describes Apollo 2 whole-body control, including walking, crouching, and stretching to manipulate objects. DeepMind contrasts this with its earlier humanoid models' upper-body tabletop tasks. These are reported capabilities and examples, not a claim of universal whole-body generalization. The five-fingered SharpaWave hand has 22 DoFdegrees of freedomThe number of independent coordinates needed to specify a mechanism's configuration. An arm's degree-of-freedom count is its number of independent joints, so a 7-DoF arm places its end effector with one coordinate to spare beyond the six a rigid pose needs, and that redundancy is what lets the elbow reconfigure while the hand stays put. More degrees of freedom buy dexterity and obstacle avoidance at the price of a larger control problem.; the Franka Duo uses two-fingered parallel grippers. The results caption identifies one model checkpoint across three configurations: Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and Franka Duo with the Robotiq gripper. That checkpoint statement does not extend to every new body adapted by On-Device 2. Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026

ER 2 plans multi-step tasks lasting several minutes, tracks progress, and coordinates with the VLA; the announcement also introduces collaboration between robots. On-Device 2 is reported to adapt to new bi-arm embodiments in a few hours of adaptation time, typically with fewer than 200 examples. The examples are Dexmate, SO101, and Trossen. The source does not provide separate example counts or evaluation protocols for each platform, and its introductory “few hours of data” wording should not be read as a measured compute-only training time. Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026

DeepMind's whole-body chart labels report 76.3% for picking from a shelf, 68.4% from a table, and 45.7% from the floor on Apollo with Inspire hands. The separate Apollo-with-Sharpa-hands chart reports individual-task results of 92% for unscrewing a bulb, 44% for tying a trash bag, 40% for Ziplock, and 32% for Dustpan. These values are explicitly written in the announcement's image alt text, not estimated from bar heights. The caption describes the whole-body bars as average success rates over multiple tasks within each skill category, while the multi-finger bars are individual-task performance. Trial counts and the meaning of the error bars are not disclosed in the captured announcement. DeepMind's conclusion is that “the multi-finger dexterous manipulation remains challenging.” Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026

Disclosure is partial, not absent. The Gemini Robotics 1.5 v3 report describes the orchestrator/VLA split and Motion Transfer at a high level, and its model card explicitly specifies continuous numerical robot actions plus text when thinking is enabled. It does not identify an action-head family, predicted/executed action-chunk lengths, a robot-command frequency, or a parameter count. Gemini Robotics Team 2025↓Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion TransferGemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan et al., 2025

The Gemini Robotics 2 announcement describes three model roles and says On-Device 2 inherits Motion Transfer from 1.5. It does not provide an internal Motion Transfer implementation, an action-output parameterization or head family, chunk lengths, a robot-command rate, or a parameter count. At that announcement, ER 2 was available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform; the VLA and On-Device models were available to early-access partners. Those access channels do not establish downloadable weights or a model-specific closed license. The inspected sources leave weight-download and licensing terms not disclosed. Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026

GR00T: the open humanoid line

The GR00T N1 report, first submitted in March 2025, describes a VLMvision-language modelA model pretrained jointly on web-scale image and text data, so visual recognition and language semantics live in one set of weights. RT-2's bet was that this pretraining is an asset for robot control: co-fine-tuning a vision-language model on robot trajectories and its original web data together transfers semantic knowledge, recognizing objects and following instructions the robot data never covered, into the policy. The vision-language-action models that followed all start from such a backbone. backbone coupled to a flow-matchingflow matchingA generative modeling recipe that learns a time-dependent vector field carrying samples from a noise distribution to the data distribution, trained by regressing the model's field against the conditional flow rather than by estimating a score. pi0 brought it to robot control: a dedicated action expert is trained with the flow-matching objective to denoise continuous action chunks, and at inference the learned field is integrated as an ODE for a handful of steps, which keeps the policy fast enough for high-frequency control while keeping the multimodality that made diffusion policies attractive. diffusion-transformer action head. Vision-language tokens condition the action head through cross-attention. The inspected v2 paper links the public GR00T-N1-2B checkpoint; it does not establish that this is the closest architectural analogue to π0 NVIDIA 2025↓GR00T N1: An Open Foundation Model for Generalist Humanoid RobotsNVIDIA, Johan Bjorck, Fernando Castañeda et al., 2025.

The captured Isaac-GR00T repository lists the N1.7 release on April 18, 2026. Its README describes a Cosmos-Reason2-2B backbone using the Qwen3-VL architecture and a 3B-parameter base checkpoint. The flow-matching DiT changes from 32 to 16 layers; the state/action interface expands to 132 dimensions and the prediction horizon to 40. NVIDIA describes a relative end-effector action space shared across robot and human embodiments, and reports 20K hours of EgoScale human video in N1.7 pretraining NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026. That README population should not be conflated with the EgoScale paper's separately reported dataset of over 20,854 hours Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. The README lists one GPU with 16 GB or more VRAM for inference; this is a hardware requirement, not a measured latency or throughput result. Prediction horizon does not specify how many actions execute per policy call NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026.

The README's License section assigns Apache 2.0 to code and NVIDIA Open Model License to model weights. Its introductory paragraph also says N1.7 is "fully commercially licensable under Apache 2.0". Those statements differ in scope; this survey records both rather than certifying an Apache license for a particular checkpoint's weights NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026.

NVIDIA describes N1.7 as comparable to N1.6, with improved generalization and language following. That wording does not establish that the backbone change was merely a platform decision or that there was no capability improvement. The sources used here are the N1 v2 technical paper and the captured N1.7 README, which links earlier N1.5 and N1.6 releases. This is a description of the inspected sources, not a claim that no later GR00T paper exists NVIDIA 2025↓GR00T N1: An Open Foundation Model for Generalist Humanoid RobotsNVIDIA, Johan Bjorck, Fernando Castañeda et al., 2025 NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026.

Figure Helix: the most detailed vendor blog

Figure’s February 20, 2025 Helix announcement describes S2, an internet-pretrained VLM, producing a continuous latent task representation, and S1, a latent-conditioned visuomotor transformer, consuming that representation together with image and robot-state inputs. Figure reports S2 at 7 to 9 Hz and S1 at 200 Hz. S1 outputs desired wrist poses, finger flexion and abduction, torso and head orientation targets, and a synthetic task-completion output, not simply a vector of upper-body joint targets. The post describes the S2 backbone as open-source and open-weight; that is not a statement that the full Helix policy was released Figure AI 2025↓Helix: A Vision-Language-Action Model for Generalist Humanoid ControlFigure AI, 2025.

Figure’s January 27, 2026 Helix 02 announcement describes a three-layer hierarchy, not three numerically specified rates. S2 interprets scenes and language and sequences latent goals. S1 maps head and palm cameras, fingertip tactile sensing, and full-body proprioception to full-body joint targets at a reported 200 Hz. S0 is a 10M-parameter whole-body controller that outputs joint-level actuator commands at a reported 1 kHz. Figure says S0 uses over 1,000 hours of joint-level retargeted human motion and is trained entirely in simulation across more than 200,000 parallel environments with domain randomization; the announcement also describes sim-to-real reinforcement learning. Figure says this learned controller replaces 109,504 lines of hand-engineered C++. S2 does not plan low-level footsteps: S1 interprets its semantic latents into motor commands and S0 executes them. The post describes its dishwasher demonstration as a continuous four-minute autonomous task with 61 loco-manipulation actions, implicit error recovery, no resets, and no human intervention. Those are Figure’s descriptions of the demonstration, not an independently timed trial or an aggregate success-rate estimate Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026.

These are Figure’s own architecture descriptions and demonstrations, not independent replications. The two linked announcement texts do not provide repeated-trial success-rate tables or release terms for the trained Helix policies. The 2025 post identifies an open-weight VLM backbone; that is not a release of the complete trained Helix policy. Read the videos as demonstrations, not success-rate estimates Figure AI 2025↓Helix: A Vision-Language-Action Model for Generalist Humanoid ControlFigure AI, 2025 Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026.

AgiBot: GO-1 and GO-2

The AgiBot World report, first submitted in March 2025, introduces GO-1. Its inspected v4 methods describe a hierarchical vision-language-latent-action (ViLLA) framework: a latent action model learns from consecutive video frames, a VLM-conditioned latent planner predicts latent action tokens, and an action expert generates low-level actions conditioned on the preceding modules. The paper explicitly includes human video without action labels and cross-embodiment robot data in training. It describes Colosseo as an open-sourced platform of data, models, benchmarks, and an ecosystem; that description does not establish that GO-1 checkpoints and the dataset were released together AgiBot-World-Contributors 2025↓AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied SystemsAgiBot-World-Contributors, Qingwen Bu, Jisong Cai et al., 2025. AgiBot's later GO-2 announcement also names ViLLA when describing GO-1 AgiBot 2026↓The Unity of Reasoning and Action: AGIBOT Unveils Genie Operator-2 (GO-2) Next-Gen Embodied Foundation ModelAgiBot, 2026.

AgiBot presents GO-2 as a response to what it calls the “Semantic-Actuation Gap” between reasoning signals and motor commands. Its Action Chain-of-Thought generates a high-level sequence of action intents as a macro-plan. In the asynchronous dual system, the lower-frequency Semantic Planning Module (System 2), called the “General Commander,” generates structured action sequences through Progressive Refinement. The higher-frequency Action Following Module (System 1), called the “Agile Executor,” combines those intents with real-time observations and uses Residual Refinement to compensate for noise. Teacher forcing is described as a training mechanism for following approximately correct but imperfect reasoning, not an extra inference-time controller. The announcement gives relative frequencies for the two modules, not numerical planner, follower, or robot-control rates AgiBot 2026↓The Unity of Reasoning and Action: AGIBOT Unveils Genie Operator-2 (GO-2) Next-Gen Embodied Foundation ModelAgiBot, 2026.

AgiBot’s announcement reports an average success rate of 98.5% across LIBERO Spatial, Object, Goal, and Long; 86.6% zero-shot success in LIBERO-Plus environments with disturbances; and, under its Genie Sim 3.0 sim-to-real heading, 82.9% real-world success for a model trained solely on simulation data. It claims superiority over π0.5 and NVIDIA GR00T, but the announcement text does not give the trial denominators, uncertainty estimates, disturbance protocol, or comparator versions needed to reconstruct a matched comparison. These are vendor-reported summaries, not interchangeable measures of frontier capability AgiBot 2026↓The Unity of Reasoning and Action: AGIBOT Unveils Genie Operator-2 (GO-2) Next-Gen Embodied Foundation ModelAgiBot, 2026. AgiBot also says the action-reasoning contribution was accepted at CVPR 2026 and the asynchronous architecture at ACL 2026. That is the vendor’s acceptance claim, not a conference-record verification. This announcement does not establish trained GO-2 weight availability; no universal “no paper” or “closed weights” conclusion follows AgiBot 2026↓The Unity of Reasoning and Action: AGIBOT Unveils Genie Operator-2 (GO-2) Next-Gen Embodied Foundation ModelAgiBot, 2026.

Skild: what the Series C announcement discloses

Skild AI's January 14, 2026 Series C announcement reports a $1.4 billion raise at a valuation of over $14 billion. The valuation is a lower bound, not an exact $14 billion. Skild calls its model “omni-bodied” and claims it can control robots without prior knowledge of their body form; this is the company's assertion, not demonstrated universal capability. Skild AI Team 2026↓Announcing Series CSkild AI Team, 2026

The announcement does contain technical descriptions: it names simulation and Internet video for pretraining, and teleoperation and deployments for post-training. Its teleoperation description maps images and proprioception to sequences of joint torques; that is a description of data, not a disclosed runtime action-head architecture. The captured announcement does not specify model architecture, parameter count, control frequency, action-chunk horizon, a reproducible benchmark protocol, or weight-download and licensing terms. That bounded disclosure gap is more precise than saying that nothing technical is publicly verifiable. Skild AI Team 2026↓Announcing Series CSkild AI Team, 2026

The field at a glance

Return to the timeline with all thirteen entries described. Selecting any node gives its capability annotation and primary source, and the π entries are cross-references to The Pi Line, included so the open and closed sides can be compared on one axis.

The split that matters for your own work is the open column. Everything downloadable as of August 2026, GR00T N1.x, GO-1, π0.5 and its predecessors, Octo, OpenVLA, sits at least one generation behind the closed frontier, and the gap is largest exactly where the vendors aim their marketing: zero-shot, open-world behavior that saturated simulation benchmarks cannot measure. The human-video transfer question is not scale alone versus a universally required alignment interface. Simar Kareer and colleagues at Physical Intelligence report transfer when π0.5 is co-fine-tuned on human data represented by 3D hand positions and relevant robot data, without a special transfer-learning mechanism; their experiments tie improved transfer to the quantity and diversity of robot pretraining Kareer 2025↓Emergence of Human to Robot Transfer in VLAsSimar Kareer, Karl Pertsch, James Darpinian et al., 2025. Ruijie Zheng and colleagues’ EgoScale combines relative wrist motion and retargeted hand actions with aligned human-robot mid-training, and reports that scale and this alignment are complementary in its dexterous-manipulation setting Zheng 2026↓EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie et al., 2026. These are different recipes and evaluations, not proof that either conclusion is universal. Cross-embodiment transfer examines their qualifications.