patchpol-so-follower-multiview

Patch Policy — real robot, two views

This private model repository contains a LeRobot-compatible export of the latest completed two-view Patch Policy checkpoint. It uses observation.images.front and observation.images.ee, with a two-frame observation context and five-action prediction horizon.

Training data: 16 episodes, 10,342 frames, 30 FPS, six joint-position actions and the so_follower robot schema. The intended training run is 50,000 optimizer steps; the latest exported snapshot in this repository is step 15,000. Raw training checkpoints are preserved under backups/.

LeRobot rollout

The root files are in LeRobot's pretrained-policy layout: config.json, model.safetensors, policy_preprocessor.json, and policy_postprocessor.json.

The custom policy must be installed once in the LeRobot environment so the standard lerobot-rollout plugin discovery can register patchpol:

uv pip install -e /path/to/this/repository/lerobot_plugin
lerobot-rollout --policy.path=atharva-pantheon/patchpol-so-follower-multiview \
  --robot.type=so100_follower

Supply the robot's normal port/calibration/camera arguments. The robot must provide both camera feature keys with the names above. DINOv2 ViT-S/14 is a frozen runtime dependency and is loaded from the local Torch Hub cache (the first run may download it).

The action postprocessor restores the recorded six-dimensional action scale using the dataset statistics included in the processor state files.