Action Chunking (ACT and ALOHA)

Action chunking is a policy design in which each inference predicts a sequence of future actions instead of one^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. The term comes from psychology, where actions grouped into one unit are cheaper to store and execute^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 ^{Source 17}17 Lai et al. 2022, Action chunking as conditional policy compression. Lucy Lai, Ann Zixiang Huang, Samuel J. Gershman, PsyArXiv, 2022.

Paper figure: Tony Z. Zhao et al. / arXiv. Licence: CC BY 4.0.

Diffusion, flow and vision-language-action policies now all chunk^{Source 18}18 Zhao et al. 2026, Dynamic Execution Horizon Prediction for Chunk-based Robot Policies. Yuchi Zhao, Miroslav Bogdanovic, Arjun Sohal et al., 2026. Try “100 moves per plan”.

Why chunking works

Behavior cloningbehavior cloningBehavior cloning is supervised learning of a policy from expert demonstrations, with observations as inputs and the expert's actions as labels. The recipe predates deep learning: ALVINN (1988) used camera and laser-range inputs and trained on simulated road images before NAVLAB tests. Scaled up, it still underlies most learned manipulation policies; its known closed-loop failure mode is covariate shift. with per-step error ε on the expert's states can accumulate cost quadratic in episode length T, because each mistake carries the robot into states the expert never visited^{Source 1}1 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011. DAggerDAggerDAgger (Dataset Aggregation) is an iterative fix for behavior cloning's distribution mismatch: the expert labels states the current policy visits, and the policy retrains on the aggregate. Ross, Gordon, and Bagnell framed it as reducing imitation learning to no-regret online learning. The task-cost guarantee is existential over the learned sequence and assumes learning and expert-mixing conditions. Linear excess-cost scaling needs a horizon-independent recovery-cost factor u, which can be O(T) at worst. relabels the states the policy visits; its excess cost is linear in T only if u, the rise in expert cost-to-go after one action, does not grow with T^{Source 1}1 Ross et al. 2011, A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. Stéphane Ross, Geoffrey J. Gordon, J. Andrew Bagnell, AISTATS 2011:

Action chunkingaction chunkingAction chunking is predicting a sequence of future actions in one inference instead of a single action per timestep. Introduced with ACT on ALOHA, it divides an episode's sequential decisions by the chunk length, reducing compounding error and giving the policy temporal context to represent pauses. The cost is commitment to a plan between replans. shortens the horizon instead: k actions per inference cut it k-fold^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. With continuous actions, smooth deterministic imitators can suffer execution error exponentially larger in the horizon than their training error, part of why chunking and diffusion policies work^{Source 19}19 Simchowitz et al. 2025, The Pitfalls of Imitation Learning when Actions are Continuous. Max Simchowitz, Daniel Pfrommer, Ali Jadbabaie, 2025, and chunking avoids that compounding through control-theoretic stability^{Source 20}20 Zhang et al. 2025, Action Chunking and Exploratory Data Collection Yield Exponential Improvements in Behavior Cloning for Continuous Control. Thomas T. Zhang, Daniel Pfrommer, Chaoyi Pan et al., 2025.

ACT: chunking with a CVAE

Action Chunking with Transformers (ACT), introduced with the ALOHA bimanual teleoperation hardware, wraps chunking in a conditional variational autoencoderCVAEA CVAE (conditional variational autoencoder) is a generative model learning a distribution over a latent variable instead of one deterministic output. ACT trains its action-chunking policy as one, so a single policy covers varied demonstrations of a task. In training, an encoder compresses each demonstrated action sequence into a latent style variable. At test time the decoder generates the chunk from camera images and joint positions, with the latent at the prior mean.: training encodes each demonstrated action sequence as a latent style variable, and at test time the policy decodes a chunk from camera images and joint positions with that latent at zero. Dropping the CVAE objective barely changes scripted-data results and cuts human-data success from 35.3% to 2%^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023.

Paper figure: Tony Z. Zhao et al. / arXiv. Licence: CC BY 4.0.

The 14 action coordinates are six joints and one gripper coordinate for each of two six-DoF ViperX arms; the paper's "7+7=14 DoF" counts the grippers. ACT's implementation description specifies L1 reconstruction, while the printed algorithm uses MSE; both add a weighted KL term^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023.

The chunk-size dial

In ACT's ablation, averaged over two simulated tasks with temporal ensembling off, success rises from 1% at k=1 to 44% at k=100 and dips at k=200 and k=400, near open-loop control; chunked BC-ConvMLP and VINN follow the same trend^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. A small k brings back compounding errorcompounding errorCompounding error is the accumulation of small per-step mistakes in a sequential task. Each error moves the robot slightly off its training data, where the next is likelier, so total deviation outgrows the per-step rate. Ross, Gordon, and Bagnell bounded it for imitation learning at total cost quadratic in episode length. Action chunking reduces it by cutting sequential decisions., and a large one reacts slowly to unexpected states^{Source 21}21 Liu et al. 2024, Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling. Yuejiang Liu, Jubayer Ibn Hamid, Annie Xie et al., 2024. The best execution horizon is task-dependent: in flow-based VLAs success rises and then falls as it lengthens^{Source 22}22 Wang et al. 2026, VLA Knows Its Limits: Adaptive Execution Horizons for Robot Policies. Haoxuan Wang, Gengyu Zhang, Yan Yan et al., 2026, and PACE, which picks it online, lifted real ALOHA and Franka success from 50.7% to 70.4%^{Source 23}23 Nie et al. 2026, PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking. Junnan Nie, Jiayi Li, Chenghao Liu et al., 2026.

Temporal ensembling and its limits

The pinned ACT reference implementation uses temporal ensemblingtemporal ensemblingTemporal ensembling is ACT's execution scheme for overlapping action chunks: the policy runs every timestep and averages in-flight predictions for the current step with exponential weights. In the pinned ACT reference implementation, predictions ordered oldest to newest get weight exp(-0.01 * i), so the oldest retained prediction has the largest weight. Handoffs are smoother, at the cost of one inference per step instead of per chunk.: it queries the policy every timestep and averages, oldest to newest, the retained predictions for the current action. Weights are $w_i = \exp(-m \cdot i)$, where $i = 0$ is the earliest prediction and $m = 0.01$, so the oldest prediction weighs most^{Source 3}3 tonyzhaozh 2023, ACT reference implementation: imitate_episodes.py (76cf30b). tonyzhaozh, 2023. The best decay rate varies by task^{Source 21}21 Liu et al. 2024, Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling. Yuejiang Liu, Jubayer Ibn Hamid, Annie Xie et al., 2024.

# 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 oldest-to-newest predictions [0, 10, 20] after selection, $m = 0.01$ gives normalized weights of about [0.33667217, 0.33332222, 0.33000561] and an average of 9.93333444; the reference occupancy filter would discard a literal zero^{Source 3}3 tonyzhaozh 2023, ACT reference implementation: imitate_episodes.py (76cf30b). tonyzhaozh, 2023.

When consecutive chunks diverge in strategy, say left versus right around an obstacle, averaging them can backfire^{Source 21}21 Liu et al. 2024, Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling. Yuejiang Liu, Jubayer Ibn Hamid, Annie Xie et al., 2024. Demonstration pauses, unpredictable from a single Markovian state, freeze single-step policies, as both ACT and Diffusion Policy document^{Source 2}2 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 ^{Source 5}5 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023.

Real-Time Chunkingreal-time chunkingReal-time chunking is an inference-time scheme keeping chunked policies responsive when inference is slower than control. While the robot finishes the current chunk, the policy conditions the next on the actions that will already have executed when inference returns, holding that prefix fixed. Black, Galliker and Levine introduced it for flow-matching policies; it tracks moving targets under delays that break naive chunking and temporal ensembling. (RTC) treats the hand-off between chunks as inpainting: it freezes actions that will already have run and generates the rest. With π0.5 at five denoising steps, model latency was 76 ms for the baselines and 97 ms for RTC, plus 10 to 20 ms over LAN. At 100 or 200 ms of injected delay, temporal-ensembling variants triggered protective stops, while RTC held average task throughput through +200 ms without retraining^{Source 6}6 Black et al. 2025, Real-Time Execution of Action Chunking Flow Policies. Kevin Black, Manuel Y. Galliker, Sergey Levine, 2025, over six tasks, ten episodes each^{Source 7}7 Black et al. 2025, Real-Time Action Chunking with Large Models. Kevin Black, Manuel Y. Galliker, Sergey Levine, 2025.

A training-time variant matched inference-time RTC at lower compute^{Source 24}24 Black et al. 2025, Training-Time Action Conditioning for Efficient Real-Time Chunking. Kevin Black, Allen Z. Ren, Michael Equi et al., 2025, LeRobot ships RTC for Pi0, Pi0.5 and SmolVLA^{Source 25}25 Hugging Face n.d., Real-Time Chunking (RTC). Hugging Face, LeRobot documentation, n.d.; accessed 2026-10-04, and A2C2's per-step correction head beats RTC by 23 points on Kinetix^{Source 26}26 Sendai et al. 2025, Leave No Observation Behind: Real-time Correction for VLA Action Chunks. Kohei Sendai, Maxime Alvarez, Tatsuya Matsushima et al., 2025. In the toy below, press “None” to watch the two plans agree.

The lineage in numbers

In the table, not disclosed means the source omits the value, and weights means download availability. Physical Intelligence's openpi README lists base checkpoints for π0, π0-FAST and π0.5^{Source 10}10 Physical Intelligence 2026, openpi. Physical Intelligence, 2026.

Octo's ALOHA fine-tuning example predicts 64 actions and executes 12 before replanning^{Source 8}8 Octo Model Team et al. 2024, Octo: An Open-Source Generalist Robot Policy. Octo Model Team, Dibya Ghosh, Homer Walke et al., 2024, and RT-1 predicts one action per timestep; its paper announces open code^{Source 11}11 Brohan et al. 2022, RT-1: Robotics Transformer for Real-World Control at Scale. Anthony Brohan, Noah Brown, Justice Carbajal et al., 2022. π0 generates its chunks with conditional flow matchingflow matchingFlow matching is a generative modeling recipe that learns a time-dependent vector field carrying noise to data, by regressing against the conditional flow instead of estimating a score. In pi0, which brought it to robot control, an action expert trained this way denoises continuous action chunks. Integrating the field as an ODE for a handful of steps is fast enough for high-frequency control and keeps diffusion policies' multimodality.^{Source 9}9 Black et al. 2024, π0: A Vision-Language-Action Flow Model for General Robot Control. Kevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024, and π0.5 combines FAST tokens with flow matching^{Source 12}12 Physical Intelligence et al. 2025, π0.5: a Vision-Language-Action Model with Open-World Generalization. Physical Intelligence, Kevin Black, Noah Brown et al., 2025, though openpi supports only its flow-matching head^{Source 10}10 Physical Intelligence 2026, openpi. Physical Intelligence, 2026.

The π0.6 model card, dated November 17, 2025, reports 63 ms per action chunk with five denoising steps and three cameras on one H100^{Source 15}15 Physical Intelligence 2025, π0.6 Model Card. Physical Intelligence, 2025. π0.7 predicts 50-step chunks and executes 15 or 25, at 20 Hz on UR5e and 50 Hz elsewhere; it distills experience from RL-trained π*0.6^{Source 16}16 Ai et al. 2026, π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities. Bo Ai, Ali Amin, Raichelle Aniceto et al., 2026. NVIDIA's GR00T N1.7 README lists a predicted horizon of 40 and downloadable weights, under Apache 2.0 in its overview and the NVIDIA Open Model License in its License section^{Source 13}13 NVIDIA 2026, NVIDIA Isaac GR00T. NVIDIA, 2026; GR00T N1 processed actions "in chunks as in Zhao et al. (2023)" with H=16^{Source 27}27 NVIDIA et al. 2025, GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. NVIDIA, Johan Bjorck, Fernando Castañeda et al., 2025.

Figure's January 27, 2026 announcement has Helix 02's S1 sending full-body joint targets at 200 Hz, tracked by S0 actuator commands at 1 kHz^{Source 14}14 Figure AI 2026, Introducing Helix 02: Full-Body Autonomy. Figure AI, 2026. Gemini Robotics turns about 250 ms of end-to-end latency into an effective 50 Hz control frequency by packing several actions into each chunk^{Source 28}28 Gemini Robotics Team et al. 2025, Gemini Robotics: Bringing AI into the Physical World. Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie et al., 2025.

Chunking on mobile robots and in RL

Mobile ALOHA moves the recipe to a wheeled whole-body platform, where co-training on static ALOHA data raises success substantially^{Source 4}4 Fu et al. 2024, Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation. Zipeng Fu, Tony Z. Zhao, Chelsea Finn, 2024, and ALOHA Unleashed scaled it to over 26,000 real demonstrations for 5 tasks with a diffusion policy^{Source 29}29 Zhao et al. 2024, ALOHA Unleashed: A Simple Recipe for Robot Dexterity. Tony Z. Zhao, Jonathan Tompson, Danny Driess et al., 2024. In RL fine-tuning, Q-chunking runs RL in a chunked action space for better exploration and unbiased n-step backups^{Source 30}30 Li et al. 2025, Reinforcement Learning with Action Chunking. Qiyang Li, Zhiyuan Zhou, Sergey Levine, 2025.