Predicting and executing a sequence of actions without intermediate replanning, known as action chunking, is increasingly used in robot learning from human demonstrations. However, its effects on learned policies remain puzzling: some studies highlight its importance for achieving strong performance, while others observe detrimental effects. In this paper, we first dissect the role of action chunking by analyzing the divergence between the learner and the demonstrator. We find that longer action chunks enable a policy to better capture temporal dependencies by taking into account more past states and actions within the chunk. However, this advantage comes at the cost of exacerbating errors in stochastic environments due to fewer observations of recent states. To address this, we propose Bidirectional Decoding (BID), a test-time inference algorithm that bridges action chunking with closed-loop operations. BID samples multiple predictions at each time step and searches for the optimal one based on two criteria: (i) backward coherence, which favors samples aligned with previous decisions, (ii) forward contrast, which favors samples close to outputs of a stronger policy and distant from those of a weaker policy. By coupling decisions within and across action chunks, BID enhances temporal consistency over extended sequences while enabling adaptive replanning in stochastic environments. Experimental results show that BID substantially outperforms conventional closed-loop operations of two state-of-the-art generative policies across seven simulation benchmarks and two real-world tasks.

In action chunking, the agent predicts the joint distribution of a sequence of actions conditioned on the context and then executes all or part of the sequence without replanning. In our analysis, we consider two models with equal context length \( c \) and investigate the effect of using a longer action horizon. Consider a shorter action chunk of length \( h \) and a longer one of length \( h + d \). The longer action chunk benefits from remembering more past states and suffers from having not observed the more recent states, as is illustrated below:

The expected loss of the two agents, \(\pi_h\) and \(\pi_{h+d}\), with respect to the expert, \( \pi^* \), are related by the following inequality. $$ \alpha_f - \epsilon_b(1 - P_b^{2d}) \leq \min_{\pi_{h+d}} \mathbb{E}_{G} \left[ \mathcal{L}(\pi_{h+d}, \pi^*) | C \right] - \min_{\pi_{h}} \mathbb{E}_{G} \left[\mathcal{L}(\pi_{h}, \pi^*) | C \right] \leq - \alpha_b + \epsilon_f(1 - P_f^{2d}). $$ Intuitively, the advantage of each policy stems from the additional information it has access to (i.e., \( \alpha_f \) for \( \pi_h \) and \( \alpha_b \) for \( \pi_{h+d} \)) while the disadvantage is bounded by the maximum divergence arising from inferring missing information incorrectly (i.e., \(\epsilon_b(1 - P_b^{2d})\) and \(\epsilon_f(1 - P_f^{2d}\)). Our theoretical analysis provides us two important takeaways :

Our hypothesis suggests that while the probability of any pair of samples sharing the same latent strategy is low, the likelihood of finding a consistent pair from a large number of samples is significantly higher. This motivates us to solve the closed-loop action chunking problem by identifying the optimal action within a batch of plans at each time step, $$ a^* = \arg \min_{a \in \mathcal{A}} \mathcal{L}_B(a) + \mathcal{L}_F(a) $$ where \( \mathcal{L}_B \) and \( \mathcal{L}_F \) are two criteria measuring the temporal dependency with respect to the backward decision and forward plan. To ensure temporal coherence, we reference the action chunk from the previous time step, \( \{ \hat{a}_{t-1}, \cdots, \hat{a}_{t+h-1} \} \) and minimize the weighted sum of Euclidean distance across \( h - 1 \) overlapping steps: $$ \mathcal{L}_B = \sum_{\tau=0}^{h-1} \rho^\tau \left\| a_{t+\tau} - {\hat a}_{t+\tau} \right\|_2. $$ This encourages consistent latent strategies across time steps, while allowing for gradual adaptation to unforeseen environment dynamics.
A robust policy should predict far enough to capture temporal dependencies in demonstrations. To ensure this, we compare each candidate plan with two reference sets: one from a well-trained model as the stronger policy, and another from an early underfitting checkpoint or a shorter prediction horizon model as the weaker policy. The forward objective minimizes the average distance to positive samples from the strong policy and maximizes the average distance to negative samples from the weak policy: $$ \mathcal{L}_F = \frac{1}{N}\left(\sum_{a^{+} \in \mathcal{A}^{+}} \sum_{\tau=0}^{l} \left\| a^{(t)}_{t+\tau} - a^{+}_{t+\tau} \right\|_2 - \sum_{a^{-} \in \mathcal{A}^{-}} \sum_{\tau=0}^{l} \left\| a^{(t)}_{t+\tau} - a^{-}_{t+\tau} \right\|_2\right),$$ where \( \mathcal{A}^{+} = \mathcal{A} \setminus \{a\} \) is the positive set predicted by the strong policy \( \pi \), \( \mathcal{A}^{-} \) is the negative set predicted by the weak policy \( \pi' \).

We evaluate BID on the Push-T, RoboMimic, and 4-Object Franka Kitchen tasks. While existing inference methods offer some benefits for closed-loop operations, they either lack robustness or are highly sensitive to decay rate. BID consistently achieves substantial gains across all tasks, surpassing the vanilla baseline by over 26% in relative improvements.