sumitagrawal/pi0-robotwin-phaseB2-30k

π0 RoboTwin — Phase B2 (30k steps)

Fine-tuned π0 policy for RoboTwin-style dual-arm manipulation, continued from lerobot/pi0_base.

Field	Value
Training run	pi0-phaseB2-20260505-151435
Checkpoint	030000
Weights	model.safetensors (+ preprocessor shards)
License	Apache-2.0

What’s in this repo

• model.safetensors — π0 policy weights (LeRobot / PI0Policy compatible).
• *.safetensors preprocessor / tokenizer artifacts — required for from_pretrained.

Quickstart (LeRobot)

Install a matching LeRobot stack (e.g. lerobot with π0 extras per LeRobot docs).

from lerobot.policies.pi0.modeling_pi0 import PI0Policy

policy = PI0Policy.from_pretrained("sumitagrawal/pi0-robotwin-phaseB2-30k")
policy.eval()
# Wire preprocessor/postprocessor from policy.config + pretrained_path for real obs batches.

For a fuller example (dataset row → batch → one forward), see the inference_pi0_from_hub.py pattern in the training project that produced this checkpoint: build an observation batch consistent with RoboTwin / ALOHA-style keys (observation.images.*, observation.state, language/task fields as expected by the preprocessor).

Training context

• Base model: lerobot/pi0_base
• Finetuning: RoboTwin-unified style data; run id and step above identify the exported pretrained_model directory used for this Hub snapshot.
• Open-loop / eval: Optional eval artifacts can be added to the card later (metrics, plots); this release is weights + minimal documentation.

Limitations

• Intended for the same observation / action convention as the training setup (14-D ALOHA-style action chunking, multi-camera specs as in training).
• Not a general-purpose VLA; validate on your robot / sim before deployment.

Citation

If you use this model, cite π0 / LeRobot and RoboTwin per their respective papers and licenses.

Model tree

Base: lerobot/pi0_base → this checkpoint (fine-tuned).