Skip to content

PiPER + Pika on mock devices

Status: in development

In development None of the PiPER, Pika Gripper and Pika Sense drivers has run on hardware. These steps were checked with mock devices only. Do not reuse the values (tool_offset and others) on hardware.

  • Goal: drive two PiPER arms with two Pika Grippers through mock Pika Senses, record, and run a policy trained with LeRobot.
  • Prerequisites: a clone of the roboloom repository (.roboloom/piper-mock/ exists only in the clone). No hardware.
  • Time: 30 minutes (not counting training).
  • Result: recordings in .roboloom/piper-mock/datasets/piper-mock, and the trained policy's runs saved in piper-mock-policy.
flowchart LR
  S["mock Pika Sense ×2"] --> K["pose_ik ×2"]
  K --> F["PiPER ×2<br/>Pika Gripper ×2"]
  F --> R["recording<br/>piper-mock"]
  R --> T["lerobot-train"]
  T --> P["infer piper-mock"]

1. Set up the environment

git clone https://github.com/tetsugo02/roboloom.git
cd roboloom
uv sync --extra piper --extra pika --extra record --extra policy --extra train

The configuration files are described in PiPER controllers.

2. Run it in the operator screen

uv run roboloom run piper-mock
  1. Start with Launch (Ctrl-S).
  2. Use "Move home" to bring the arms to poses/<arm>/ready.yaml. The arms start folded (all joints 0), where the IK cannot follow a hand.
  3. A mock Sense engages 2 s after opening, then repeats 10 s engaged and 2 s disengaged. The controller rows turn following.
  4. Start recording with Start episode; Stop and Save store it.

3. Record episodes of a fixed length

uv run roboloom record piper-mock --task "pick" --seconds 15 --episodes 5 --home --plain
uv run roboloom check-dataset piper-mock

--home moves the arms to the start pose before every episode.

4. Train with LeRobot

uv run lerobot-train \
  --dataset.repo_id=local/piper-mock \
  --dataset.root=.roboloom/piper-mock/datasets/piper-mock \
  --policy.type=act --policy.device=cpu --policy.push_to_hub=false \
  --trackio.enable=false --steps=200 --batch_size=8 \
  --output_dir=outputs/train/act_piper_mock

This only checks the path, so it trains briefly on a CPU. Before using a GPU, check with labstatus that one is free.

5. Run a policy

First run the demo policy in the same directory (demo_policy.py).

uv run roboloom infer piper-mock --seconds 10 --plain

Then run the trained checkpoint and save the runs for evaluation with --task.

uv run roboloom infer piper-mock --seconds 20 --task "pick" \
  --policy roboloom.policy.lerobot:load \
  --policy-kwargs "{\"path\": \"$PWD/outputs/train/act_piper_mock/checkpoints/last/pretrained_model\"}"
uv run roboloom check-dataset piper-mock --with policy
  • The inference configuration has no Senses, so a policy cannot use observation.left_sense and the like.
  • Interventions (teleop:) are not available: pose_ik does not support them.

Next