ACT on LIBERO goal task 0 (open the middle drawer of the cabinet)
Single-task ACT policy trained with LeRobot on one LIBERO task.
Task
open the middle drawer of the cabinet in LIBERO libero_goal, task index 0
(task_index=19 in lerobot/libero).
Training
| Dataset | lerobot/libero, 43 episodes of this task only |
| Steps | 100,000 |
| Batch size | 8 |
| Params | 51.6M |
| Hardware | 1× RTX 5070 Ti (16 GB) |
| Wall clock | 57 min |
Evaluation
lerobot-eval, libero_goal task 0, seed 1000, 10 episodes, n_action_steps=10:
| Checkpoint | Success rate | VRAM peak | Sec/episode |
|---|---|---|---|
| 20k steps | 60.0 % | 1283 MiB | 3.6 |
| 100k steps (this) | 100.0 % | 1283 MiB | 2.7 |
Reference points measured under the same protocol on the same task:
| Policy | Params | Success rate | VRAM peak |
|---|---|---|---|
| ACT (this) | 0.05 B | 100.0 % | 1283 MiB |
| SmolVLA | 0.5 B | 100.0 % | 1975 MiB |
| MolmoAct2 | 5 B | 90.0 % | 13541 MiB |
Scope and limitations
ACT is not language-conditioned, so this checkpoint only covers the task it was
trained on. Evaluated zero-shot on a neighbouring task (libero_goal task 1) it
scores 0.0 % (0/10). Use it for this drawer task only; train a separate
policy per task.
Usage
lerobot-eval \
--policy.path=AmberHyunKIM/act_libero_goal_task0 \
--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.