Comparison Matrix

RT-2’s reported rates depend on the model and serving setup. Its 55B PaLI-X variant runs at 1 to 3 Hz, while the 5B variant runs at around 5 Hz, using a multi-TPU cloud service queried over the network. Brohan 2023↓RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic ControlAnthony Brohan, Noah Brown, Justice Carbajal et al., 2023

OpenVLA v3 reports approximately 6 Hz inference on one NVIDIA RTX 4090 in bfloat16, without compilation, speculative decoding, or other inference speed-up tricks. Kim 2024↓OpenVLA: An Open-Source Vision-Language-Action ModelMoo Jin Kim, Karl Pertsch, Siddharth Karamcheti et al., 2024

The eight axes

Action representation describes the output mechanism: continuous regression, discrete action tokens, or a generative head such as diffusion or 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.. Training supervision and runtime action generation need not use the same representation.

Horizon separates the planned chunk length $H$ from the steps executed $\hat{H}$ before re-planning: the predicted horizonpredicted horizonThe number of future actions a chunked policy emits in one inference, the chunk length H. The policy predicts that many actions in a single forward pass whether or not it will execute them all: π0 predicts 50 actions, one second at 50 Hz, and Diffusion Policy predicts 16. What a reader of a results table needs alongside that number is how many of the predicted actions the system actually runs, which is the committed horizon. and the committed horizoncommitted horizonThe number of predicted actions a system actually 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 already in flight, which is what lets the committed horizon stay shorter than the predicted one without a pause between chunks.. Chi 2023↓Diffusion Policy: Visuomotor Policy Learning via Action DiffusionCheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023 The Diffusion Policy CNN Push-T configuration predicts 16 steps and executes 8; these are not universal settings for every architecture or task. Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026 π0.7 predicts 50 and executes either 15 or 25, not every integer between them. Its paper does not map those two execution choices to particular robots.

Control frequency needs a named measurement setting. Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 ACT's ALOHA setup runs at 50 Hz; with temporal ensemblingtemporal ensemblingACT's execution scheme for overlapping action chunks: the policy is queried at every timestep, and the predictions that each in-flight chunk made for the current step are averaged with exponential weights. In the pinned ACT reference implementation, retained predictions are ordered oldest-to-newest and weighted by exp(-0.01 * i), so the oldest retained prediction has the largest weight. This smooths the handoff between chunks at the cost of one inference per step instead of one per chunk., the policy is queried at every timestep. Brohan 2022↓RT-1: Robotics Transformer for Real-World Control at ScaleAnthony Brohan, Noah Brown, Justice Carbajal et al., 2022 RT-1 commands the Everyday Robots platform at 3 Hz. Neither statement is a standalone transformer-throughput benchmark.

Backbone, conditioning, and cross-embodiment describe the model components, their inputs, and the scope of robot-body transfer. Distinguish mixed-embodiment training, direct use of a checkpoint on multiple bodies, and finetuning on a new body rather than treating them as equivalent evidence of transfer.

Hierarchy distinguishes an internal high-level stage from a separately paired system. Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025 π0.5 predicts a semantic subtask before generating low-level actions. NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026 In the N1.7 SONIC workflow, a separate whole-body controller decodes the VLA's latent action tokens into joint commands. Here, external does not mean that the controller is a reasoning layer above the policy. Weights records download availability, not license openness; the notes identify the checked source scope.

Brohan 2022↓RT-1: Robotics Transformer for Real-World Control at ScaleAnthony Brohan, Noah Brown, Justice Carbajal et al., 2022 RT-1’s “limited” cross-embodiment label refers to mixing Kuka IIWA data with Everyday Robots data, then evaluating on Everyday Robots. It does not mean that one checkpoint was demonstrated controlling both robot bodies.

Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 ACT’s 100-step entry names the paper’s k=100 chunk-ablation reference; temporal ensembling queries the policy every timestep. The paper trains a roughly 80M-parameter model from scratch for each task and announces software and hardware. That does not establish pretrained-checkpoint downloads or one checkpoint transferring across embodiments; those cells are source-scoped as not disclosed.

Chi 2023↓Diffusion Policy: Visuomotor Policy Learning via Action DiffusionCheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023 Diffusion Policy’s CNN Push-T reference uses observation/prediction/execution horizons of 2/16/8. Its real-world UR5 Push-T experiment predicts commands at 10 Hz and interpolates them to 125 Hz. These are named settings, not defaults for every task or backbone. The checked paper announces code, data and training details, not a pretrained-checkpoint catalogue; no shared-checkpoint cross-embodiment result is established here.

Brohan 2023↓RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic ControlAnthony Brohan, Noah Brown, Justice Carbajal et al., 2023 RT-2’s multi-TPU cloud service runs PaLI-X-55B at 1 to 3 Hz and the 5B version at around 5 Hz. The paper also studies PaLM-E-12B and a separate Language-Table model. “Limited” denotes these separate setup-specific models, not demonstrated cross-body use of one checkpoint. The action-only table entry excludes the separately evaluated chain-of-thought prompting variant. The inspected paper does not establish model-specific weight-release terms.

Kim 2024↓OpenVLA: An Open-Source Vision-Language-Action ModelMoo Jin Kim, Karl Pertsch, Siddharth Karamcheti et al., 2024 OpenVLA v3 reports approximately 6 Hz inference on one RTX 4090 in bfloat16, without compilation, speculative decoding or other speed-up tricks. Its Franka-Tabletop and Franka-DROID fine-tuning setups separately use 5 Hz and 15 Hz non-blocking controllers. The paper describes single-step action prediction, multi-robot pretrained use and fine-tuning for new setups, and explicitly releases downloadable checkpoints. Its 7B label names the language-model scale; the visual encoder is separately reported as 600M parameters.

Kim 2025↓Fine-Tuning Vision-Language-Action Models: Optimizing Speed and SuccessMoo Jin Kim, Chelsea Finn, Percy Liang, 2025 The OFT row uses the paper's ALOHA OFT+ setting: 25 predicted and executed actions at a 25 Hz controller. LIBERO instead predicts and executes eight. Its limited cross-embodiment label means adaptation from the pretrained model to separate tasks and robot setups, not demonstrated control of every body by one unchanged checkpoint.

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 Gemini Robotics 1.5's v3 model card does disclose continuous numerical action outputs. It does not identify the action-head family, action-chunk length, robot-command rate, or parameter count. Internal refers to thinking within the VLA; the report also pairs it with a separate GR-ER 1.5 orchestrator. The generalization and cross-embodiment experiments use one checkpoint across ALOHA, bi-arm Franka, and Apollo; the separate long-horizon agent experiment adds post-training on bi-arm Franka. An unset field here is source-scoped, not inapplicable.

Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026 The Gemini Robotics 2 row describes the VLA in DeepMind's July 30, 2026 announcement. External hierarchy means the separate ER 2 model coordinates with that VLA; it is not an undisclosed internal action head. The same-checkpoint result covers Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and Franka Duo with Robotiq, not every On-Device adaptation. The announcement does not specify the VLA backbone, output parameterization, chunk lengths, control frequency, or parameter count. AI Studio, private preview, and early-access partnerships are access channels, not weight-license classifications.

Skild AI Team 2026↓Announcing Series CSkild AI Team, 2026 Skild's 2026 row is scoped to its January 14 Series C announcement, not a verified model-release date. The company describes its data sources and makes an omni-bodied capability claim, but does not specify a reproducible cross-embodiment evaluation or the runtime architecture, inputs, hierarchy, chunk lengths, control frequency, and weight-release terms used by these cells. The empty conditioning list means not disclosed in this source, not no inputs or n/a.

Helix 02 and GO-2 disclosure limits

Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026 Figure’s Helix 02 announcement describes S1 joint targets at 200 Hz and S0 actuator commands at 1 kHz, but does not specify the output-head training loss. The representation cell is therefore not classified as continuous regression here. Its statement about generalization across Figure’s fleet is not evidence of one checkpoint transferring between different robot bodies; that field remains not disclosed for this source.

AgiBot 2026↓The Unity of Reasoning and Action: AGIBOT Unveils Genie Operator-2 (GO-2) Next-Gen Embodied Foundation ModelAgiBot, 2026 AgiBot describes lower-frequency System 2 semantic planning and higher-frequency System 1 action following, not two numerical control rates. The inspected announcement text does not establish the backbone, output encoding, action-chunk lengths, cross-body checkpoint transfer, or trained-policy weight-release terms. Those fields remain not disclosed, rather than n/a or closed. The 2026 year is supported by an HTML publication timestamp in a commented span; it is not an independently established release date.

What the matrix can establish

Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 ACT fixes its CVAE latent to the prior mean, zero, to decode deterministically at test time. Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024 π0 instead uses an action expert trained with flow matching to generate continuous action chunks. These are concrete differences between output mechanisms, not evidence that every later policy abandoned regression.

Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025 The π0.5 paper distinguishes FAST-token training from continuous-action generation with flow matching. It describes autoregressive action decoding as expensive at inference time. That qualitative comparison is not a measured factor-of-two task-completion or inference-latency result for π0-FAST.

Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026 π0.7 uses five denoising steps to generate 50 actions, then executes either 15 or 25. Its UR5e robots run at 20 Hz; the other tested robots run at 50 Hz. Black 2025↓Real-Time Execution of Action Chunking Flow PoliciesKevin Black, Manuel Y. Galliker, Sergey Levine, 2025 Real-time chunking addresses inference delay by guiding generation of the next chunk while the current one is executing. These execution details matter more than ranking methods by predicted horizon alone.

Physical Intelligence 2025↓π0.6 Model CardPhysical Intelligence, 2025 The π0.6 card reports 63 ms to produce an action chunk with five denoising steps, three camera inputs, and one H100. It does not disclose a robot-control rate or a predicted/executed chunk length. Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026 Figure reports Helix 02 S1 joint targets at 200 Hz and S0 actuator commands at 1 kHz. The latter are different stages of a control stack, not two alternative speeds for one model call.

Brohan 2022↓RT-1: Robotics Transformer for Real-World Control at ScaleAnthony Brohan, Noah Brown, Justice Carbajal et al., 2022 The RT-1 paper announces code, which does not establish pretrained-weight availability. Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026 Figure's Helix 02 announcement does not state a chunk horizon or weight-release terms. Their availability cells therefore read "not disclosed" for these checked sources. Physical Intelligence 2026↓openpiPhysical Intelligence, 2026 The pinned openpi README lists π0, π0-FAST, and π0.5 checkpoints. NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026 The N1.7 README lists downloadable weights, while its License section distinguishes Apache 2.0 code from the NVIDIA Open Model License for weights. These release statements do not establish a global generation ranking or closed licensing for an unlisted model.

The Control Hz column is not an inference-throughput comparison. Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 ACT's 50 Hz ALOHA loop, Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026 Helix's 200 Hz S1/1 kHz S0 stack, and Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026 π0.7's 20 Hz UR5e/50 Hz other-robot settings describe different systems. π0.7 also trains with simulated inference delays; that does not turn these rates into a shared benchmark. Read the setting notes before comparing two cells.

Use the matrix as a constraint solver

Start with elimination, not ranking. Remove models whose weights or training data cannot be used under the project's terms. Remove models whose action representation cannot map cleanly to the robot's controller, then those whose measured end-to-end latency misses the deadline. Next test observation compatibility and the amount of adaptation data the model requires. Only after those constraints should benchmark performance influence the shortlist.

For the surviving candidates, run one controlled comparison with the same cameras, action horizon, reset policy and physical trial protocol. Keep a simple behavior-cloning or action-chunking baseline in the evaluation. The foundation-model guide explains what prior each family contributes; the action-space guide explains whether its output is a fair fit. A matrix can expose missing evidence, but it cannot turn results from different robots and protocols into a universal leaderboard.