Action Chunking (ACT and ALOHA)

Behavior cloningbehavior cloningTraining a policy by supervised learning on expert demonstrations: each recorded observation is an input, the action the expert took at that moment is the label, and the fitted mapping from state to action is the policy. The recipe predates deep learning; ALVINN (1988) used camera and laser-range inputs and trained on simulated road images before NAVLAB tests, and the same approach, scaled to modern networks and datasets, still underlies most learned manipulation policies. Its known failure mode in closed loop is covariate shift. fails in closed loop for a precise reason: a policy with per-step error ε on the expert's state distribution can accumulate cost that grows quadratically with episode length, because each mistake carries the robot into states the expert never visited Ross 2011↓A Reduction of Imitation Learning and Structured Prediction to No-Regret Online LearningStéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011. Over a T-step episode the worst-case cost scales as

where the second bound is what DAggerDAggerDataset Aggregation, the iterative fix for behavior cloning's distribution mismatch: roll out the current policy, have the expert label the states the policy actually visits, add those labeled states to the training set, and retrain. Ross, Gordon, and Bagnell framed the procedure as a reduction of imitation learning to no-regret online learning, whose task-cost guarantee is existential over the learned sequence and requires learning and expert-mixing assumptions. Linear excess-cost scaling also requires the expert recovery-cost factor u to be independent of the horizon; u can be O(T) in the worst case. buys by repeatedly relabeling the states the policy actually visits Ross 2011↓A Reduction of Imitation Learning and Structured Prediction to No-Regret Online LearningStéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011. Action chunkingaction chunkingPredicting a sequence of future actions in one inference instead of a single action per timestep. Introduced with ACT on the ALOHA platform, chunking cuts the number of sequential decisions in an episode by the chunk length, which reduces compounding error and gives the policy the temporal context it needs to represent pauses, at the cost of committing to a plan between replans. attacks the same bound from a different side: it shrinks the effective horizon. Instead of predicting one action per inference, the policy predicts a sequence of k actions at once, so the number of sequential decisions in an episode drops by a factor of k.

Drag the chunk size slider to k=1 and then back to 100. That curve is ACT's own ablation: 1% success one action at a time, 44% at a hundred. Everything else in this module explains why those two numbers are so far apart.

ACT: chunking with a CVAE

Action Chunking with Transformers (ACT), introduced together with the ALOHA bimanual teleoperation hardware, wraps chunking in a conditional variational autoencoderCVAEConditional variational autoencoder: a generative model that learns a distribution over a latent variable instead of a single deterministic output. ACT trains its action-chunking policy as a CVAE. During training an encoder compresses the demonstrated action sequence into a latent style variable; at test time the decoder generates the chunk conditioned on camera images and joint positions with the latent clamped to the prior mean, which lets one policy cover varied demonstrations of the same task. Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. During training, an encoder reads the demonstration's action sequence and produces a latent style variable. At test time the policy decodes a chunk conditioned on camera images, joint positions, and the latent clamped to zero.

The 14 action coordinates are six arm joints and one gripper coordinate per arm. The ACT hardware section names two six-DoF ViperX arms, while Section IV-C uses the shorthand “7+7=14 DoF”; that notation must not be read as seven arm joints Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. Mobile ALOHA explicitly describes its 14 arm-action coordinates as including two continuous gripper actions Fu 2024↓Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body TeleoperationZipeng Fu, Tony Z. Zhao, Chelsea Finn, 2024.

The objective in the card follows ACT's implementation description: Section IV-C specifies L1 reconstruction, but Algorithm 1 prints MSE, with a weighted KL regularizer in both accounts Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. The paper is internally inconsistent here; the card states the Section IV-C choice, not a universal ACT objective.

The chunk-size dial

Chunk size is a real bias/variance dial, not a hyperparameter footnote. The curve above comes from the paper's two simulated MuJoCo tasks (Cube Transfer and Bimanual Insertion), averaged over four settings (each task trained on scripted and on human demonstrations, temporal ensembling disabled): ACT measures 1% success at k=1, 44% at k=100, and a decline at k=200 and k=400 as the policy becomes effectively open-loop Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. Too small and compounding errorcompounding errorThe accumulation of small per-step mistakes over the course of a sequential task: each error carries the robot into a state slightly outside its training data, where the next error is more likely, so total deviation grows faster than the per-step error rate. Ross, Gordon, and Bagnell bounded the effect for imitation learning with total cost quadratic in episode length, and action chunking attacks the same effect from the other side by shrinking the number of sequential decisions in an episode. returns; too large and the policy stops reacting to the world mid-chunk.

Temporal ensembling and its limits

The pinned ACT reference implementation uses 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.: it queries the policy at every timestep and averages the retained predictions for the current action in oldest-to-newest order tonyzhaozh 2023↓ACT reference implementation: imitate_episodes.py (76cf30b)tonyzhaozh, 2023. Its weights are $w_i = \exp(-m \cdot i)$, where $i = 0$ is the earliest retained prediction and $m = 0.01$. Thus the oldest retained prediction has the largest weight; increasing $m$ reduces the relative weight of newer predictions. The example below starts with already-selected predictions: it illustrates the weighted average, not the reference code's nonzero-buffer selection rule.

Original schematic

Diagram: Robot Wiki contributors / Robot Wiki (original diagram). Licence: CC BY 4.0.

# ACT reference convention: preds contains current-step predictions already
# selected and ordered oldest-to-newest. Index 0 is the oldest retained
# prediction, not "zero chunks ago". Reference m=0.01. This helper does not
# reproduce ACT's raw-buffer nonzero occupancy filter.
def temporal_ensemble(preds: list[np.ndarray], m: float = 0.01) -> np.ndarray:
    weights = np.exp(-m * np.arange(len(preds)))   # w_i = exp(-m * i)
    weights /= weights.sum()
    return np.tensordot(weights, np.stack(preds), axes=1)

For an illustrative scalar example after selection, oldest-to-newest predictions [0, 10, 20] at $m = 0.01$ receive normalized weights approximately [0.33667217, 0.33332222, 0.33000561] and yield 9.93333444. This does not exercise the reference occupancy filter, which would discard a literal zero scalar in the raw buffer.

Real-Time Chunkingreal-time chunkingAn inference-time scheme that keeps chunked policies responsive when the model is slower than the control loop. While the robot finishes the current chunk, the policy conditions its next prediction on the actions that will already have executed by the time inference returns, holding that committed prefix fixed and generating only the remainder. Black, Galliker, and Levine introduced it for flow-matching policies, where it tracks moving targets under inference delays that break both naive chunk execution and temporal ensembling. (RTC) reframes the hand-off between chunks as an inpainting problem on the flow or diffusion trajectory: actions that will already have executed are frozen, the overlapping middle is partially attended to, and the remainder is generated fresh Black 2025↓Real-Time Execution of Action Chunking Flow PoliciesKevin Black, Manuel Y. Galliker, Sergey Levine, 2025. In the reported π0.5 experiment, RTC showed no degradation in average task throughput through +200 ms of injected delay, without retraining Black 2025↓Real-Time Execution of Action Chunking Flow PoliciesKevin Black, Manuel Y. Galliker, Sergey Levine, 2025. This evaluates RTC on 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. policy; it does not establish adoption by later generations of the pi line.

The dial below is a deterministic toy of a cross-mode hand-off, not a reproduction of the RTC benchmark. Its percentages are a normalized toy score: the collapse curve, the fixed RTC line, and the interpolation between measured delay settings are assumptions. The paper reports average task throughput, not these percentages; the slider also extends beyond the experiment’s +200 ms endpoint Black 2025↓Real-Time Execution of Action Chunking Flow PoliciesKevin Black, Manuel Y. Galliker, Sergey Levine, 2025.

The lineage in numbers

The table compares reported prediction horizons and control settings, not universal deployment defaults. Weights describe download availability, not license openness; an unverified release is not disclosed, not evidence of a closed license. Applicable unpublished horizons and rates also read not disclosed, never n/a or zero.

Physical Intelligence 2026↓openpiPhysical Intelligence, 2026 Physical Intelligence’s openpi README lists base checkpoints for π0, π0-FAST, and π0.5 with download locations, as observed on 7 September 2026. The cited repository snapshot is dated 24 August 2026; the README was last edited 21 November 2025. This catalogue is not a global latest-release claim, and absence from it does not establish another model’s license.

Octo v2's ALOHA finetuning example predicts 64 actions and executes 12 before replanning. Its other reported setups include Franka action prediction at 15 Hz, a coffee-task controller at 10 Hz, and ViperX end-effector control at 5 Hz. These are separate examples, not one universal Octo control rate or horizon. The action head uses diffusion Octo Model Team 2024↓Octo: An Open-Source Generalist Robot PolicyOcto Model Team, Dibya Ghosh, Homer Walke et al., 2024.

Black 2024↓π0: A Vision-Language-Action Flow Model for General Robot ControlKevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024 π0 v4 predicts 50 actions. In its reported execution setup, UR5e and Franka run at 20 Hz and execute 16 actions before replanning; the other evaluated robots run at 50 Hz and execute 25. The table's 20 Hz entry names the UR5e/Franka reference, while the paper describes operation at up to 50 Hz. Its continuous action outputs are trained with conditional flow matching.

Brohan 2022↓RT-1: Robotics Transformer for Real-World Control at ScaleAnthony Brohan, Noah Brown, Justice Carbajal et al., 2022 RT-1 predicts one robot action per timestep and commands the Everyday Robots mobile manipulator at 3 Hz. Each action dimension is mapped to one of 256 uniformly distributed bins within that variable’s bounds. Six input frames are observation history, not a six-action horizon. The paper announces open code; pretrained-weight availability is not disclosed in that statement.

Black 2025↓π0.5: a Vision-Language-Action Model with Open-World GeneralizationPhysical Intelligence, Kevin Black, Noah Brown et al., 2025 The π0.5 v1 paper predicts 50 action entries (inclusive indexing, H=49) and sends targets at 50 Hz in its two-mobile-platform setup. That does not establish how many actions execute before replanning. The paper combines FAST-token prediction with flow matching; the cited openpi README supports only the flow-matching head for π0.5 training and inference.

Physical Intelligence 2025↓π0.6 Model CardPhysical Intelligence, 2025 The November 17, 2025 π0.6 model card describes a continuous-action flow-matching expert; FAST tokens supervise the VLM backbone during training. The card does not specify predicted or executed chunk length, robot-control rate, or model-specific weight-release and licensing terms. Its 63 ms action-chunk latency uses five denoising steps and three camera inputs on one H100; this is inference latency, not a control frequency. The card’s openpi release statement concerns the π0.5 comparator, not π0.6.

Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026 In the π0.7 paper’s experiments, the continuous-action flow-matching expert predicts 50-step chunks with five denoising steps and executes either 15 or 25 steps. UR5e robots run at 20 Hz; the other tested robots run at 50 Hz. The paper does not assign either execution choice to a particular robot. FAST tokens supervise the backbone during training, not the runtime action output. The model distills experience from RL-trained π*0.6; evaluation without task-specific post-training does not mean training without RL-derived data. Model-specific weight-release and licensing terms are not disclosed in this paper.

NVIDIA 2026↓NVIDIA Isaac GR00TNVIDIA, 2026 The NVIDIA Isaac GR00T README describes the N1.7 GA release: a predicted horizon of 40, a flow-matching DiT head, relative-end-effector support, and downloadable weights. Relative EEF is not its only embodiment interface; it also describes SONIC latent-action output. No single robot-control frequency is disclosed there; inference throughput is a different quantity. Its overview says Apache 2.0, but its License section distinguishes Apache 2.0 code from weights under the NVIDIA Open Model License. Download availability does not resolve that licensing discrepancy.

Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026 Figure’s January 27, 2026 Helix 02 announcement reports S1 full-body joint targets at 200 Hz, tracked by S0 actuator commands at 1 kHz. These are separate layers, not interchangeable control rates. Chunk length and weight-release/licensing terms are not disclosed in this announcement; neither absence establishes inapplicability or closed licensing.

Where the idea travels next

Mobile ALOHA extends the same recipe to a wheeled whole-body platform and shows that co-training on static ALOHA data lifts mobile manipulation success substantially Fu 2024↓Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body TeleoperationZipeng Fu, Tony Z. Zhao, Chelsea Finn, 2024.