What Is Action Chunking in Robot Policies?

A policy that emits chunks is trained on examples that pair each observation with a run of consecutive actions, not a single one. For humanoid and other robot foundation models trained on egocentric human video, the two papers cited here build that run of actions from wrist motion and hand articulation, or from latent actions, rather than from a robot’s own commands.

Firsthand does not deliver robot actions or action chunks. It captures 3D hand pose, wrist trajectories, contact events and discrete action segments directly at the time of recording, in the same session as the egocentric video and depth, collected fast, on demand and to a written spec, and delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. A buyer training a chunked policy chooses the chunk length and builds the action representation on top of those labels; this site states no recommended chunk length and no chunked export.

Action chunking is a way of shaping what a robot policy outputs. Instead of predicting a single action for the current moment, the model predicts a block of consecutive future actions, called a chunk, in one pass. The length of the block is a setting chosen by the model’s authors. This page quotes the two official sources that describe chunks directly: NVIDIA’s GR00T N1 and EgoScale papers.

Chunking is a property of the policy, not of the dataset. In the GR00T N1 formulation the chunk is defined by the model at each time t. A dataset supplies observations and action labels, and the policy’s training setup decides how a run of those labels becomes a chunk. For a buyer that distinction is practical: you do not buy a chunk, you buy the labelled sequences a chunked policy is trained from.

The GR00T N1 paper (arXiv 2503.14734) introduces “an open foundation model for humanoid robots”: a vision-language-action model whose action module is a diffusion transformer trained with flow matching. On how that module handles time, the paper states: “The actions are processed in chunks as in Zhao et al. (2023).” It then defines the chunk as the action vectors of timesteps t through t+H−1, and adds: “We set H=16 in our implementation.” The related-work section describes the same design as a flow-matching action generation model “with action chunking.”

The paper also ties the chunk to inference. It states that “the inference time for sampling a chunk of 16 actions is 63.9ms on an L40 GPU using bf16,” and that at inference time it generates action chunks with K-step denoising, where “we found K=4 inference steps to work well across all embodiments.” One sampling call therefore returns 16 action vectors, not one.

GR00T N1 also shows what a chunk is built from when the source is human video. For human videos, it states, “in the absence of ground-truth actions, we extract learned latent actions and use them as flow-matching targets.” The chunk is the same output shape, while the targets come from a learned latent representation, not from a recorded wrist pose.

EgoScale (arXiv 2602.16710) pretrains a flow-based vision-language-action policy on 20,854 hours of egocentric human video, using wrist motion and retargeted dexterous hand actions as supervision. Its architecture section says: “The model then predicts a chunk of future actions using a flow-matching objective.”

The paper writes arm motion relative to the chunk. It states: “Given timestep t in an action chunk,” wrist motion is expressed as the transform from the wrist pose at the chunk’s first timestep to the pose at timestep t. That removes dependence on the absolute camera pose, so the same representation can be shared between human demonstrations and robot executions. The validation procedure uses chunks directly: at each sampled timestep “we draw 16 samples from the flow-matching policy, average the predicted action chunks, and compute the mean squared error against ground-truth wrist and hand actions.”

EgoScale is also candid about where its human action labels come from. It recovers camera motion and hand trajectories with off-the-shelf SLAM and hand-pose estimation pipelines, and says “these estimates are noisy due to unconstrained data collection.”

Both papers describe chunking as how their policies output actions, and both report results for the whole system. In the passages cited here neither paper reports an ablation of chunk length or attributes a specific gain to chunking alone, so this page does not either.

What the sources do show is the mechanics. GR00T N1 pairs a vision-language module that runs at 10 Hz with an action module that generates closed-loop motor actions at 120 Hz, and it samples a 16-action chunk in a single call. EgoScale pairs the same flow-matching output with wrist and hand targets taken from human video.

Three terms are easy to blur. An action label is the recorded or derived quantity attached to a video: a hand pose, a wrist trajectory, an action segment, a contact event. An action representation is the format the action half of a training example takes, such as a wrist-level pose or a latent action. Action chunking is how many consecutive actions, in that representation, the policy predicts at once.

In EgoScale the representation is wrist-level pose and hand joint actions, and the chunk is the group of timesteps over which it is predicted. In GR00T N1 for human video the representation is a latent action, and the chunk is again the output block.

This site states that Firsthand captures 3D hand pose, wrist trajectories, contact events and discrete action segments, each segment with a start time, end time, verb and noun, directly at the time of recording and in the same session as the egocentric video and depth. A wrist trajectory is the position and orientation of the wrist over time, which is the kind of quantity EgoScale builds its chunks from. The site does not state that Firsthand delivers robot actions, action chunks, a windowed export or its own action representation.

The collection itself is fast and on demand. The site states that brief to written spec takes the same week, and that the published clock starts at the signed spec: a median of 48 hours to the first validated batch where coverage is already deep, and three to four weeks where contributors must be recruited first in a thinner environment. Delivery is in RLDS, LeRobot format or WebDataset under a buyer-owned commercial license. If your policy is chunked, name the labels you need and your target delivery format in the spec, and keep the chunk length on your side of the pipeline.

Several things a team might assume are not stated and should be raised in the spec conversation, not inferred: a recommended chunk length; any chunked or time-windowed export; delivery of robot joint or gripper commands; validation that a particular chunk length suits a particular label type; and any result showing Firsthand data improves a chunked policy. The only action-chunking figures on this page, H = 16, the 63.9 ms sampling time and K = 4 denoising steps, belong to GR00T N1 and are not Firsthand specifications.

What is action chunking?

Action chunking is a policy design in which a robot model predicts a block of consecutive future actions at once, called a chunk, instead of one action per step. GR00T N1 describes its actions as processed in chunks of H = 16 timesteps.

How long is an action chunk?

It is a setting the authors choose. GR00T N1 states “We set H=16 in our implementation.” The EgoScale passages cited on this page do not give a chunk length, and this site states no recommended length.

Does Firsthand deliver action chunks or robot actions?

No. Firsthand captures 3D hand pose, wrist trajectories, contact events and discrete action segments directly at the time of recording. A buyer chooses the chunk length and builds the action representation in their own training pipeline.

Can a chunked policy be trained on egocentric human video?

EgoScale does so: it pretrains a flow-based policy that predicts a chunk of future actions on 20,854 hours of egocentric human video, using wrist motion and retargeted hand actions. GR00T N1 uses learned latent actions as targets for human video instead.

Is action chunking the same as an action representation?

No. The representation is the format of an action, such as a wrist-level pose or a latent action. Chunking is how many consecutive actions in that representation the policy predicts at once.

How do I get a custom dataset for a chunked policy fast?

Write the labels and delivery format you need into a skill spec. The site states brief to spec takes the same week, with a median of 48 hours to the first validated batch where coverage is deep and three to four weeks where contributors must be recruited first, both from the signed spec.