Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control?
Filippo Lazzati, Kyle Stachowicz, William Chen, Alberto Maria Metelli, Andrew Wagenmaker, Sergey Levine
cs.RO
2026-08-04
Action chunking is standard in robot imitation learning but poorly understood. This work rules out three old hypotheses and pins the real gains on predicting from past observations plus implicit ensembling; a randomized-delay ensemble matches it, an explicit ensemble beats it.
Action chunking, which has a policy predict and execute several consecutive actions at once, is a de facto staple of robot imitation learning; diffusion policy and ACT both lean on it for gains. But why it works has been mostly folklore, never rigorously tested. This paper supplies the explanation.
The authors systematically test three common hypotheses on the Libero-90 benchmark (90 tabletop tasks), the Robomimic benchmark, and three Franka real-robot tasks, then offer their own account. Policies are diffusion policies with image observations, and every result is averaged over at least three random seeds.
The three hypotheses under test: temporal consistency (human demonstrations are smooth and correlated, which chunking models better); horizon reduction (predicting many steps cuts error accumulation); and representation learning (chunking yields better features).
The verdict: the gains of the first two are fully captured by something far simpler, and the third does not hold up.
The old hypotheses fail. On Libero-90, a "delayed policy" that predicts the current single action from the observation k steps ago matches or beats action chunking. What actually matters is the non-Markovian trick of predicting from a past observation; smoothly replaying an action sequence is unnecessary, so temporal consistency is not the point. The finding reproduces on the VLA model π0.5: run as a delayed policy, it performs as well as standard action chunking.
Implicit ensembling is the real key. On the harder Robomimic tasks, delayed policies fall behind action chunking. The authors' explanation is a third benefit: during training a chunked policy simultaneously learns aₜ|oₜ, aₜ|oₜ₋₁, …, aₜ|oₜ₋k₊₁, a pack of predictors over differently delayed features that aggregate like a random forest to cut variance. A single delayed policy uses only one of them, so it loses.
A randomized-delay ensemble (RDE) replaces chunking. If the crux is using all the temporal relationships, then sample a random delay each step and predict from there:
| Task | Single-step | Action chunking | Delayed | Randomized-delay ensemble |
| Libero-90 | 68.9 | 89.2 | 94.0 | 93.6 |
| Robomimic Tool Hang | 28.0 | 75.2 | 51.6 | 71.8 |
| Robomimic Transport | 3.3 | 12.6 | 7.9 | 12.1 |
RDE matches action chunking across every setting, including the three real-robot tasks, without ever replaying action chunks.
An explicit ensemble does better still. Instantiate the "implicit ensemble" explicitly: train m independent policies and aggregate, and you beat action chunking. Robomimic Transport climbs from 12.6% to 41.5% (about +29 points), Tool Hang from 75.2% to 87.6%.
For robotics practitioners the takeaway is hands-on. Action chunking is not magic; its entire benefit decomposes into "predict from past observations" and "ensemble," and both have cheaper or stronger substitutes. Use a randomized-delay ensemble as a drop-in replacement, or an explicit ensemble for a higher ceiling. This is a chunking-free, interpretable improvement path.
The authors concede an important boundary: control frequency. At the 15–20 Hz they use, delayed policies suffice, but at 50–60 Hz they fall behind chunking unless sub-chunks of length 5 are treated as single actions, pulling the effective frequency back to 10–20 Hz. They speculate this tracks human visually guided behavior, which sits around 2–10 Hz. Temporal consistency therefore matters again at high frequency, so the conclusions carry a frequency caveat.
The real-robot experiments have an engineering wrinkle too: pausing to compute the next action lets the robot drift, which they work around by computing the next action while the current one executes, at the cost of a fixed one-step delay. Whether that handling muddies the conclusions is not examined. And Transport and Tool Hang have low absolute success rates (12.6%, 75.2%): the gains look good, but "solved" is still far off.