ACT on LIBERO goal task 0, 20k steps (under-trained variant)

Single-task ACT policy trained with LeRobot on one LIBERO task.

This is the under-trained checkpoint, published for a training-budget comparison. For the working policy use AmberHyunKIM/act_libero_goal_task0 (100k steps, 100 % success).

Task

open the middle drawer of the cabinet in LIBERO libero_goal, task index 0 (task_index=19 in lerobot/libero).

Training

Identical to the 100k run except for --steps.

Why this checkpoint exists

Training steps were the only variable changed between the two runs, so the pair isolates how much of ACT's performance on this task comes purely from training budget.

lerobot-eval, libero_goal task 0, seed 1000, 10 episodes, n_action_steps=10:

Both peak at 1283 MiB VRAM. The 5× longer run is both more accurate and faster per episode, because a better policy finishes the task in fewer steps.

Scope and limitations

ACT is not language-conditioned, so this checkpoint only covers the task it was trained on, and at 60 % it fails this task four times in ten. Treat it as an ablation artifact, not a policy to deploy.

Usage

lerobot-eval \
  --policy.path=AmberHyunKIM/act_libero_goal_task0_20k \
  --policy.device=cuda \
  --policy.n_action_steps=10 \
  --env.type=libero --env.task=libero_goal --env.task_ids='[0]' \
  --eval.batch_size=1 --eval.n_episodes=10 --seed=1000

Requires MUJOCO_GL=egl on a headless machine.