Comparison Matrix

The preceding modules in this domain each followed one lineage at a time. This one lays them side by side. The table below covers 18 policies, from RT-1 in 2022 to the current frontier, across the eight axes that determine how a learned policy behaves in a control loop. Every value comes from a primary source: a paper, a model card, or the vendor's own documentation, and the row carries that source. Two honesty rules apply throughout. Cells the vendor has not published read "not disclosed" instead of a guessed value, and figures that circulate without verification, such as the RT-2 and OpenVLA control rates, are omitted rather than repeated Brohan 2023RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic ControlAnthony Brohan, Noah Brown, Justice Carbajal, 2023 Kim 2024OpenVLA: An Open-Source Vision-Language-Action ModelMoo Jin Kim, Karl Pertsch, Siddharth Karamcheti et al., 2024.

The eight axes

Action representation is how the policy emits motor commands: continuous regression, discrete tokens decoded like text, or a generative head (diffusion or flow matching) that samples continuous chunks. This choice drives everything downstream, including inference cost and control rate.

Horizon shows the planned chunk length $H$ against the steps actually executed $\hat{H}$ before the policy re-plans. The gap between the two numbers sets the policy's reactivity: Diffusion Policy predicts 16 steps but commits 8, and π0.7 predicts 50 but commits only 15 to 25 Chi 2023Diffusion Policy: Visuomotor Policy Learning via Action DiffusionCheng Chi, Zhenjia Xu, Siyuan Feng et al., RSS 2023, 2023 Physical Intelligence 2026π0.7: a Steerable Model with Emergent CapabilitiesPhysical Intelligence, 2026.

Control frequency is the policy-side rate, not the low-level controller's. ACT queries its transformer at 50 Hz and smooths with temporal ensembling Zhao 2023Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023, 2023; RT-1 managed 3 Hz Brohan 2022RT-1: Robotics Transformer for Real-World Control at ScaleAnthony Brohan, Noah Brown, Justice Carbajal, 2022. Dual-rate stacks report both rates in the cell note.

Backbone, conditioning, and cross-embodiment describe what the model is built on, what signals it consumes, and whether one checkpoint serves multiple robot bodies. Cross-embodiment "yes" hides three different mechanisms, padded joint vectors in the π line, a shared relative end-effector frame in GR00T N1.7, and DeepMind's unnamed Motion Transfer recipe Black 2024π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024 NVIDIA 2026Isaac GR00T (N1.7 release notes and code)NVIDIA, 2026 Gemini Robotics Team 2025Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion TransferGemini Robotics Team, Google DeepMind, 2025.

Hierarchy marks whether a slow reasoning layer sits above the fast policy, and whether that layer lives inside one model (internal, as in π0.5's subtask prediction) or in a separate system (external, as when GR00T pairs with a whole-body controller) Black 2025π0.5: a Vision-Language-Action Model with Open-World GeneralizationKevin Black, Noah Brown, James Darpinian et al., CoRL 2025, 2025. Weights records whether you can download the model today.

What the matrix shows

Generative action heads displaced regression. ACT's deterministic CVAE decoder is the last entry in the continuous-regression column; from late 2023 onward every row samples actions through diffusion, discrete tokens, or flow matching Chi 2023Diffusion Policy: Visuomotor Policy Learning via Action DiffusionCheng Chi, Zhenjia Xu, Siyuan Feng et al., RSS 2023, 2023 Black 2024π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024. The discrete-token detour of RT-1, RT-2, and OpenVLA bought web-scale pretraining at the cost of autoregressive decode latency, and the field's answer was the flow-matching action expert: continuous chunks in a handful of denoising steps, trained behind a stop-gradient so the VLM backbone keeps its web knowledge Driess 2025Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize BetterDanny Driess, Jost Tobias Springenberg, Brian Ichter et al., NeurIPS 2025, 2025. π0-FAST is the exception that tests the rule: it keeps the autoregressive recipe and makes it fast enough with DCT tokenization, at roughly twice the end-to-end latency of the flow head Pertsch 2025FAST: Efficient Action Tokenization for Vision-Language-Action ModelsKarl Pertsch, Kyle Stachowicz, Brian Ichter et al., RSS 2025, 2025.

Chunks got long, then execution got selective. Horizon went from 1 step in RT-1 to 50 in the π line, but the newest rows shrink the committed horizon rather than the planned one. π0.7 executes 15 to 25 of its 50 predicted steps, and the real-time chunking work exists precisely to keep long-chunk policies responsive under inference latency Physical Intelligence 2026π0.7: a Steerable Model with Emergent CapabilitiesPhysical Intelligence, 2026 Black 2025Real-Time Execution of Action Chunking Flow PoliciesKevin Black, Manuel Y. Galliker, Sergey Levine, 2025.

Fifty hertz became the standard claim, and the note column shows what it costs. Temporal ensembling, parallel decoding, KV-cached prefixes, and five-step denoising all exist to make a multi-billion-parameter model answer inside a 20 ms budget Kim 2025Fine-Tuning Vision-Language-Action Models: Optimizing Speed and SuccessMoo Jin Kim, Chelsea Finn, Percy Liang, 2025 Physical Intelligence 2025π0.6 Model CardPhysical Intelligence, 2025. Helix 02's split is the extreme version: no single rate closes the loop, so a 200 Hz visuomotor policy hands off to a 1 kHz whole-body controller Figure AI 2026Introducing Helix 02: Full-Body AutonomyFigure AI, 2026.

Hierarchy moved inside the model. From π0.5 onward the frontier rows all carry an internal high-level layer: subtask prediction in the π line, the ER orchestrator in Gemini Robotics, the S0/S1/S2 stack in Helix Black 2025π0.5: a Vision-Language-Action Model with Open-World GeneralizationKevin Black, Noah Brown, James Darpinian et al., CoRL 2025, 2025 Parada 2026Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, Google DeepMind, 2026. The external-hierarchy era of separate planners is covered in the next module.

Finally, the weights column tells the access story in one glance. Everything downloadable, ACT through OpenVLA-OFT, π0 through π0.5, GR00T N1.7, sits at least one generation behind the closed frontier, and the closed rows are exactly the ones with the most "not disclosed" cells Black 2025π0.5: a Vision-Language-Action Model with Open-World GeneralizationKevin Black, Noah Brown, James Darpinian et al., CoRL 2025, 2025 Bjorck 2025GR00T N1: An Open Foundation Model for Generalist Humanoid RobotsJohan Bjorck, Yuke Zhu, NVIDIA, 2025.

One comparison the table cannot make is frequency against frequency. ACT's 50 Hz, Helix's 200 Hz S1, and π0.7's 50 Hz with training-time latency compensation are three different measurement setups, not three points on one scale. Treat the Control Hz column as "the rate the source reports for the policy" and read the cell notes before ranking two rows against each other.