Vanh Tran
Physical AI · 2 min read

What chunk_size does to an ACT policy

A single-variable LeRobot ACT experiment on PushT: success falls from 14% to 8% to 0% as chunk_size grows from 16 to 32 to 100.

I trained ACT policies on the PushT task and changed one thing: chunk_size, the number of future actions the policy predicts at once. Each policy was evaluated on 50 episodes.

Results

chunk_sizeSuccessAvg. max rewardAvg. total reward
1614% (7/50)0.68671.29
328% (4/50)0.70679.15
1000% (0/50)0.32628.28

Success counts an episode only when the T-block holds its coverage of the goal, so a brief graze doesn't count.

The finding

Success falls every time the chunk grows, but the best single moment peaks at 32. With 32-step chunks the policy approaches the goal more smoothly and then overshoots. It commits to 32 actions without looking again, and the drift builds up before the next re-plan. At 100 it re-plans only about three times an episode and never gets close.

The bottleneck is open-loop staleness, not the training horizon.

The full write-up, per-episode data and the steps to reproduce it are in the original post and the repo.

Read the full write-up

Written by Vanh Tran

Builder from Hanoi, building Geovily in public. More about me