AmberHyunKIM/dagger_libero10_task6

DAgger aggregated dataset: MolmoAct2 teacher on LIBERO-10 task 6

Training data used to distill MolmoAct2 (5B) into SmolVLA (0.45B) on a single LIBERO task, assembled with DAgger (Ross et al., 2011).

• Task: libero_10 task 6, "put the white mug on the plate and put the chocolate pudding to the right of the plate" (long-horizon, two objects in sequence)
• 66 episodes / 19,691 frames, 10 fps, 256x256, two cameras

What is in it

Two parts, aggregated as the DAgger paper prescribes (D <- D u D_i).

Part	Episodes	Collected by	Actions labelled by
Teacher trajectories	36	MolmoAct2 driving the robot	MolmoAct2
Student-visited states	30 (12 failures)	the distilled SmolVLA student	MolmoAct2 (relabelled)

The second part is the point of DAgger. Teacher success trajectories never contain the states where the student actually goes wrong, so a student trained only on them never learns to recover. Here the student was run in the environment, every episode it visited was kept including failures, and the teacher then answered "what would I do here?" for each of those states.

Teacher rollouts were filtered to successes only (36 saved out of 38 attempts, 94.7 %).

Features

Key	Type	Shape
observation.images.image	video	256 x 256 x 3 (agentview)
observation.images.image2	video	256 x 256 x 3 (wrist)
observation.state	float32	(8,)
action	float32	(7,)

action is the teacher's action after the environment post-processor, i.e. the 7-D command actually sent to the simulator.

What was measured with it

lerobot-eval, same task, seed 1000, 50 episodes, n_action_steps=10.

Model	Trained on	Success
SmolVLA baseline	human demos, all 40 LIBERO tasks	48.0 %
SmolVLA distilled	teacher part only (8,409 frames)	56.0 %
SmolVLA DAgger	this dataset (19,691 frames)	58.0 %
SmolVLA full-VLM	this dataset, VLM unfrozen	58.0 %
MolmoAct2 (teacher)		100 %

Aggregating 2.3x more data, aimed exactly at the student's failure modes, moved 56 % to 58 %. Unfreezing the VLM on top of that moved it by 0 pp. An off-the-shelf X-VLA (0.9B) scores 92 % on this task without any of this.

Published so the negative result is reproducible, not because the resulting policies are good.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("AmberHyunKIM/dagger_libero10_task6", video_backend="pyav")
print(ds.meta.total_episodes, len(ds))

video_backend="pyav" matters on machines without FFmpeg shared libraries, where torchcodec fails to load.

Caveats

• Single task, one seed range. Collection used seeds 2000+, evaluation seed 1000, so train and eval initial states do not overlap, but this is still one task.
• Only the teacher's final 7-D action is recorded. No output distribution, no intermediate features. Anyone wanting deeper distillation signal must regenerate.
• Actions come from MolmoAct2 running in bf16 + AMP.
• Episode boundaries between the two parts are not tagged in the dataset itself. The first 36 episodes are teacher trajectories, the remaining 30 are relabelled student states.