Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control?

Abstract

Action chunking—predicting and executing multiple actions instead of a single action—has proven to be a critical component for learning effective robotic control policies. However, our precise understanding of why action chunking improves performance has remained limited. In this work we seek to close this gap. Through rigorous experimental evaluations in both simulated and real-world settings, we show that existing hypotheses for the success of action chunking—temporal consistency, horizon reduction, and representation learning—fail to explain the success of action chunking. Instead, we find that action chunking benefits from greater non-Markovian expressivity and reduced compounding error compared to Markovian policies, but, in many settings of interest, these effects can be fully captured by delayed policies, which at each step predict a single action based on the observation $k$ steps in the past. We then show that there exists an additional benefit of action chunking that we refer to as implicit ensembling. In particular, by learning a diversity of temporal relationships (that is, $a_t | o_t, a_t | o_{t-1}, \ldots$), action-chunked policies exhibit behavior matching that of a model ensemble, increasing their robustness and generalization ability over policies that only learn a single temporal relationship. Building on these insights, we show that in simulated and real-world robotic control settings, we can match the performance of action chunking without action chunking—by deploying an action chunking policy as an ensemble of policies with randomized delays. Furthermore, we propose a policy class that amplifies the benefits of action chunking by explicitly instantiating an ensemble, and which we show significantly improves over the performance of action chunking in many domains.

Mechanism 1

Action chunking improves performance by capturing non-Markovianity in human demonstrators, but "temporal consistency" is not, in general, required, and it suffices to simply predict actions based on past observations

Human demonstrations can contain non-Markovian behavior that a stationary Markovian policy may struggle to fit, but the temporal consistency instantiated by action chunking can more effectively capture. We find that this indeed partially explains the success of action chunking, but temporal consistency is not required—instead, the key component is predicting actions based on past states. On LIBERO-90, we find that delayed policies—that predict a new action at each step, but pased on a past observation, $a_t \mid o_{t-n} $—can match or exceed the performance of action chunking.

LIBERO validation loss plot — Action prediction loss on held-out LIBERO-90 demonstration trajectories. Demonstrator actions can be more accurately predicted from an observation 10 steps in the past than the current observation.

LIBERO success rate plot — Success rate on LIBERO-90 of action chunking policies and delayed policies, for different action chunk length and delay $n$. Delayed policies exceed the performance of action chunking for the correct choice of delay.

Mechanism 2

Action chunking reduces compounding error by predicting actions based on past observations

The poor performance of behavioral cloning can often be attributed to compounding errors: a small mistake moves the policy away from the demonstration distribution, making later mistakes more likely. In the worst case, a Markovian behavioral cloning policy can suffer exponentially with the horizon:

Theoretical result

Lower bound for Markovian behavioral cloning

There exists a deterministic environment with horizon $H$, a Markov demonstrator $\pi_{\mathrm{demo}}$, and a Markov policy $\hat{\pi}$ with supervised learning loss $\epsilon$, such that $\mathcal{J}(\pi_{\mathrm{demo}})\ge \mathcal{J}(\hat{\pi})+\Omega(2^H \cdot \epsilon)$.

We find that action chunking can improve on this; an action chunking policy with chunk size $k$ scales exponentially with $H/k$ instead of $H$. This is due, however, not to the fact that action chunking reduces the "effective horizon" by only predicting every $k$ steps, but because, on average, the action chunking policy predicts actions based on states farther in the past, and states farther in the past are likely to have compounded less error. Given this, we find that the delayed policy predicting $a_t \mid o_{t-n}$ at each step can achieve the same guarantee:

Theoretical result

Upper bound for action chunking and delayed behavioral cloning

Suppose that the generalization error of each action chunk component/delayed policy is no larger than that of a Markovian learner, and is bounded by $\epsilon$. Then both the action-chunked and delayed policies satisfy $\mathcal{J}(\pi_{\rm demo})-\mathcal{J}(\hat\pi_k),\; \mathcal{J}(\pi_{\rm demo})-\mathcal{J}(\pi_{\rm delay}^k[\hat\pi_k]) \le \mathcal{O}((k+1)^{H/k} \cdot \epsilon)$, where $k$ is the chunk size/delay.

We also show this is tight; action chunking cannot improve on this scaling. Thus, a second benefit of action chunking is its ability to reduce compounding error by predicting actions based on past states, but this is again captured by delayed policies.

Mechanism 3

Action chunking implicitly instantiates an ensemble and benefits from ensemble-like generalization

While our first two insights suggest that delayed policies often capture the beneficial effects of action chunking, we find that this is not always the case—in some settings action chunking is still necessary. To explain this, our key insight is that an action chunking policy learns many temporal relationships at once: $a_t \mid o_t$, $a_t \mid o_{t-1}$, and so on, learning to predict a single action $a_t$ from many different "features". This is, essentially, how many ensemble-based approaches (e.g. random forests) operate—training an ensemble of models to predict the same target based on different features. Thus, action chunking can be seen as implicitly instantiating an ensemble of delayed policies.

We deploy an action chunking policy as a "randomized delay ensemble": at each step, sample an index $i \sim \mathrm{unif}([k])$, then compute and execute $[\hat{\pi}_k(o_{k-i+1})]_i$. In other words, at each step we select a random delay, predict the action chunk based on the observation at this delay, then execute the action from this chunk corresponding to the current step. In this way we both predict actions based on past observations, but also leverage all learned temporal relationships present in an action chunking policy—capturing all three effects without deploying actual action chunks. We see that across simulated settings (LIBERO and Robomimic), this essentially matches the performance of action chunking—action chunking is not necessary.

Real-World Evidence

Randomized Delay Policies Match the Performance of Action Chunking in Real-World Robotic Control

We test whether these insights hold in real-world settings. We evaluate on a Franka Emika robot arm with three manipulation tasks: putting a carrot in a bowl, removing bread from a toaster, and putting sushi in a cup. Each task uses 50 demonstrations, and each policy is evaluated for 50 rollouts per task. We consider four deployment approaches: action chunking, single-step, delayed, and randomized delay ensemble.

Real-world task performance — Real-world success rates across three manipulation tasks.

We see that delayed policies improve substantially over the single-step baseline, showing that conditioning on past observations captures an important part of the benefit. Randomized delayed execution then matches, and slightly improves over, standard action chunking on average. Just as we see in simulation, this suggests that the real-world improvements due to action chunking do not require true action chunk executation—predicting actions based on past states, and leveraging the ensemble-like effects of action chunking fully captures the performance of action chunking.

Four methods side by side for each task. Use the rollout buttons below each task to switch examples.

BibTeX

@article{lazzati2026chunking,
  author  = {Filippo Lazzati and Kyle Stachowicz and William Chen and Alberto Maria Metelli and Andrew Wagenmaker and Sergey Levine},
  title   = {Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control?},
  year    = {2026},
}