Algorithm 1: Adaptive Action Chunking Policy

Input: Set: Trained policy parameters $\theta$; allowable chunk set $K$; max steps $T_{\mathrm{m}\mathrm{a}\mathrm{x}}$.

Initialize: Obtain initial observation $o_0$, set $t=0$.

While task not successful and do

Encode: Encode current multi-view obs. $o_t$: $f_t=E_{\theta }\left(o_t\right)$.

Predict in Parallel: Process with two heads: Action Head: Predict action seq. $A_t=[a_t,\dots a_{t+K_{\mathrm{m}\mathrm{a}\mathrm{x}}-1}]$. Chunk Head: Predict distribution $P\left(K_t\right|o_t)$, set $K_t=\arg \max P\left(K_t\right|o_t)$, where $K_t\in K$.

Action Selection: Slice the first actions from $A_t$: $A_t=[a_t,\dots a_{t+K_t-1}]$.

Open-loop Execution: Execute the actions on the robot without replanning.

Update: $t\leftarrow t+K_t$. Obtain new observation $o_t$ after execution.

End While