Manipulation & Learned Policies
Other Generalist Policies
Gemini Robotics, GR00T, Helix, Skild, and GO-2: the closed-model landscape and how to read vendor claims.
Last reviewed 2026-08-08
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?
Gemini Robotics: the strongest closed line, the thinnest disclosure
Gemini Robotics 1.0 (March 2025) is a Gemini 2.0-based VLA shipped alongside Gemini Robotics-ER, an embodied-reasoning model specialized for spatial understanding, pointing, grasp proposal, and trajectory prediction Gemini Robotics Team 2025. The 1.5 release (October 2025) claims three innovations: the VLA interleaves natural-language thinking traces with actions, a training recipe called Motion Transfer carries behavior across embodiments, and ER 1.5 acts as a high-level orchestrator with a tunable thinking budget Gemini Robotics Team 2025. Both releases have technical reports on arXiv, which puts them a tier above most of this module in provenance.
Gemini Robotics 2 (July 2026) is three models. The VLA controls an Apptronik Apollo 2 humanoid from feet to fingertips, dynamically managing center of gravity so the robot steps, squats, and bends rather than reaching from a static base, and it drives a 22-DoF five-fingered SharpaWave hand as well as parallel grippers on a Franka Duo. One checkpoint covers all three embodiment pairs. ER 2 is the high-level agent, planning multi-minute tasks and coordinating multiple robots. On-Device 2 runs locally and adapts to a new bi-arm embodiment in a few hours with typically under 200 examples Parada 2026.
The reported numbers are worth quoting because DeepMind quotes them honestly. Whole-body picking on Apollo 2: 76.3% from a shelf, 68.4% from a table, 45.7% from the floor. Multi-finger dexterity with the SharpaWave hand: 92% unscrewing a bulb, but 44% tying a trash bag, 40% closing a ziplock, 32% with a dustpan. Their own framing: multi-finger dexterous manipulation remains challenging Parada 2026.
What you cannot get is the architecture. Neither the 1.5 report nor the 2 announcement discloses the action representation, chunk size, control frequency, action head type, or parameter counts. Motion Transfer is named in two releases and specified in neither. Weights are closed throughout; ER is reachable through an API, the VLA and On-Device variants through early-access partnerships Gemini Robotics Team 2025 Parada 2026.
GR00T: the open humanoid line
GR00T N1 (March 2025) is NVIDIA's open foundation model for generalist humanoids: a VLM backbone for perception and language, with a flow-matching diffusion-transformer action head that denoises continuous actions, the two coupled by cross-attention Bjorck 2025. It is the closest open analogue to π0's action-expert design, and it shipped with weights.
The current release, N1.7 (April 2026), swaps the backbone to Cosmos-Reason2-2B (a Qwen3-VL derivative) at 3B parameters total, keeps a 16-layer flow-matching DiT head with state and action dimension 132 and an action horizon of 40, and moves to a shared relative end-effector action space. That last choice is what lets 20K hours of egocentric human video (EgoScale) enter pretraining directly: human hands and robot grippers live in the same frame. Inference needs a single 16 GB GPU. Code is Apache-2.0, weights are under NVIDIA's Open Model License NVIDIA 2026.
Two honesty notes. First, the README itself says N1.7 delivers comparable performance to N1.6 with improved generalization and language following, so the backbone swap was a platform decision, not a capability jump. Second, provenance: only N1 has a paper. N1.5, N1.6, and N1.7 are documented through repository release notes, which is thinner than the π line's PDFs even though the code and weights are open Bjorck 2025 NVIDIA 2026.
Figure Helix: the most detailed vendor blog
Helix (February 2025) introduced the System 1 / System 2 split for Figure's humanoids: S2, a VLM, reasons slowly about scenes and instructions and emits latent goal vectors; S1, a fast visuomotor transformer conditioned on those latents, produces upper-body joint targets from pixels Figure AI 2025.
Helix 02 (January 2026) adds System 0, a learned whole-body controller, and the stack now runs at three rates. S2 sequences behaviors. S1 runs at 200 Hz and is all-sensors-in, all-joints-out: head and palm cameras, fingertip tactile sensors, and full-body proprioception in, complete joint-level targets for legs, torso, head, arms, wrists, and fingers out. S0 is a 10M-parameter network running at 1 kHz, trained on over 1,000 hours of retargeted human motion plus sim-to-real RL across 200,000 parallel environments, and Figure says it replaced 109,504 lines of hand-engineered C++. The architectural point of S0 is that it learns a single prior over how humans move while balancing, so the layers above never reason about footsteps. The headline demo is a 4-minute continuous dishwasher load and unload, 61 loco-manipulation actions with implicit error recovery and no human intervention Figure AI 2026.
All of this is vendor-reported. There is no paper, no weights, no published success rates, and no baselines; the source is a company blog post with videos. The architecture description is specific and plausible, but nothing here has been independently checked Figure AI 2026.
AgiBot: one open, one closed
GO-1 (March 2025) is built on ViLLA, vision-language-latent-action. A latent action token sits between the VLM and the low-level action head, and because latent actions can be inferred from unlabeled video, human or robot, the model absorbs video that carries no action labels. GO-1 was open-sourced alongside the AgiBot World Colosseo platform AgiBot Research 2025.
GO-2 (April 2026) targets what AgiBot calls the semantic-actuation gap: in a standard VLA the reasoning signal and the motor commands drift apart, and errors accumulate over long horizons. The fix has two parts. An action chain-of-thought first generates a sequence of high-level action intents as a macro plan, then executes it stage by stage. And an asynchronous dual system pairs a low-frequency planner (the "general commander") with a high-frequency follower (the "agile executor") that refines the intents against real-time observations, trained with teacher forcing so it tolerates imperfect reasoning AgiBot 2026.
The reported results are vendor numbers: LIBERO average 98.5%, LIBERO-Plus 86.6% zero-shot under disturbances, and 82.9% real-world success when trained on Genie Sim 3.0 simulation data alone, with claims of outperforming π0.5 and GR00T Report 2026. Read the LIBERO figure with the saturation caveat that the evaluation module develops: at 98.5% average, LIBERO is not meaningful evidence of frontier capability. GO-2 has no paper, and unlike GO-1 its weights are closed AgiBot 2026.
Skild: a funding announcement, not a method
Skild AI claims an "omni-bodied" general-purpose brain, one foundation model unifying perception, planning, and control across arbitrary robot bodies. The claim arrived alongside a $1.4B Series C in January 2026 at a valuation above $14B Skild AI 2026. What is publicly verifiable is essentially nothing technical: no paper, no weights, no benchmark results, no architecture disclosure. Skild belongs in the atlas as a landscape entry and a valuation signal, not as a method you can evaluate. If that changes, this section will too.
The field at a glance
The timeline below lays every release on one axis. Color encodes access: amber nodes are downloadable, dim nodes are closed. Shape encodes provenance: circles have papers, squares have repo release notes, triangles are lab blogs, diamonds are press releases. Filter to open or closed with the buttons; select any node for its capability annotation and primary source. The π entries are cross-references to The Pi Line, included so the open and closed sides can be compared on one axis.
Helix Figure Feb 2025 closed weights
provenance: lab blog, vendor-reported
System 1 / System 2 split: a VLM reasons slowly and emits latent goals; a fast visuomotor transformer turns pixels into upper-body joint targets.
Source: Figure AI 2025 (vendor-reported)
Vendor-reported is not verified
Seven of the thirteen releases above rest on lab blogs or press releases with no external replication: both Helix generations, both closed π generations, Gemini Robotics 2, GO-2, and Skild. Some of those claims will hold up. The point is that a triangle or diamond on the timeline is a company describing itself, and the honest way to cite those numbers is with the vendor's name attached, never as independent results.
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 comparison-matrix module puts these systems side by side across the eight architectural axes, and the cross-embodiment module picks up the live disagreement this survey only touched: whether human video needs an explicit shared action space, as GR00T N1.7 assumes, or whether transfer emerges from scale alone, as Physical Intelligence reports.