Modelpublic

khanhnd61/octo-small_so101-multi-task-clean

sourceHugging Faceapache-2.0updated 12d agoView on Hugging Face

0likes59downloads

Model Card

Model Card for octo

This is a octo policy trained with LeRobot.

This policy has been trained and pushed to the Hub using LeRobot.

See the full LeRobot documentation.

What changed in this revision

Retrained at Octo-Small's documented recipe (batch_size=32, steps=20000, per src/lerobot/policies/octo/README.md) with image augmentation enabled, replacing a previous batch_size=16 / 9500-step run that was an ACT-style budget and left this model underfit (9.9 epochs against the documented 41.8).

Measured on 6 held-out episodes (2 per task, eval_split=0.1, never seen in training):

	previous revision	this revision
Open-loop MAE on held-out episodes	2.40 deg	2.35 deg
Within-chunk jerk (deg/step^2)	1.178	0.341

The jerk figure is the one that matters here. Octo re-plans every 4 frames (7.5 Hz at 30 FPS) and its diffusion head resamples each time, so sampling variance reaches the servos as visible shaking. This revision is 3.5x smoother while being slightly more accurate, and is smoother than an ACT-based policy trained on the same data (0.455).

Two notes for anyone reproducing this:

—*The last checkpoint is the right one, and held-out loss says otherwise.* Diffusion eval loss was lowest around step 6000 (0.538) and worst at step 20000 (0.952), yet step 20000 is dramatically the best on actual sampled actions - step 4000 produces a jerk of 7.29. Average denoising error is a poor proxy for sample quality here; score the actions.
—`n_inference_samples` defaults to 8 in this checkpoint. Octo's reference implementation draws one sample per chunk; averaging 8 cuts sampling jitter about 3x and costs almost nothing, because the transformer runs once and only the small score network repeats (13.7 -> 13.5 ms per re-plan). Override with --policy.n_inference_samples=1 to reproduce upstream exactly.

Known data defect

The follower arm's wrist_roll never tracked the leader during recording: the action column spans 111-163 deg while the recorded state sits frozen at ~87.5 deg (per-episode std 0.00, action/state correlation about 0). Every policy trained on this dataset commands wrist_roll 60-70 deg beyond what the arm can reach, on 100% of timesteps. Inert for task behaviour, but it drives that servo into its stop for the whole rollout.

Model Details

—License: apache-2.0
—Robot type: so_follower
—Cameras: front, wrist

Inputs & Outputs

The policy consumes these observation features and produces these action features.

Inputs

Feature	Type	Shape
`observation.images.front`	VISUAL	`(3, 256, 256)`
`observation.images.wrist`	VISUAL	`(3, 128, 128)`

Outputs

Feature	Type	Shape
`action`	ACTION	`(6,)`

Training Dataset

—Repository: khanhnd61/so101-multi-task-clean
—Episodes: 44
—Frames: 15317
—Frame rate: 30 FPS
—Task(s): "Put the tape into the box", "Put the tape into the cup", "Put the cup into the box"

Training Configuration

Setting	Value
Training steps	20000
Batch size	32
Optimizer	adamw
Learning rate	0.0003
Seed	1000
LeRobot version	0.6.1

How to Get Started with the Model

New to LeRobot? These guides cover the full workflow:

—[Install LeRobot](https://huggingface.co/docs/lerobot/main/en/installation) — set up the lerobot package.
—[Hardware setup](https://huggingface.co/docs/lerobot/main/en/hardware_guide) — assemble, wire, and calibrate your robot and cameras.
—[Record data & train a policy](https://huggingface.co/docs/lerobot/en/il_robots) — the end-to-end imitation-learning walkthrough.
—[CLI cheat-sheet](https://huggingface.co/docs/lerobot/main/en/cheat-sheet) — quick reference for the lerobot-* commands.

The short version to run and train this policy:

Run the policy on your robot

bash

lerobot-rollout \
  --strategy.type=base \
  --robot.type=so_follower \
  --robot.port=<your_robot_port> \
  --robot.cameras="{ <camera_1>: {type: opencv, index_or_path: <index_or_path>, width: 640, height: 480, fps: 30}, <camera_2>: {type: opencv, index_or_path: <index_or_path>, width: 640, height: 480, fps: 30}}" \
  --policy.path=khanhnd61/octo-small_so101-multi-task-clean \
  --task="Put the tape into the box" \
  --duration=60

Replace the remaining <...> placeholders with your own values: --robot.port and the camera names/indices are specific to your machine, and the camera names must match the observation keys this policy was trained on.

When --strategy.type=base is used the script doesn't record the episodes. Skipping duration will make the policy run indefinitely. For more information look at rollout documentation.

Train your own policy

bash

lerobot-train \
  --dataset.repo_id=${HF_USER}/<dataset> \
  --policy.type=octo \
  --output_dir=outputs/train/<policy_repo_id> \
  --job_name=lerobot_training \
  --policy.device=cuda \
  --policy.repo_id=${HF_USER}/<policy_repo_id> \
  --wandb.enable=true

Writes checkpoints to `outputs/train/<policyrepoid>/checkpoints/`.

Evaluation

<!-- Report real-robot results here: run the policy several times per task and count the successes. Delete the "No evaluation results" line and fill in this table instead:

Task	Trials	Successes	Success rate
pick the lego brick	10	8	80%

Also worth noting: anything that affects difficulty (new object positions, lighting, distractors, a different robot of the same type, ...). -->

No evaluation results have been provided for this policy yet.

Citation

If you use this policy, please cite the method linked in the description above, along with LeRobot:

bibtex

@misc{cadene2024lerobot,
    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
    title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
    howpublished = "\url{https://github.com/huggingface/lerobot}",
    year = {2024}
}