Skip to content

Running your own policy

Mock only

  • Goal: train the sample policy chunk_mlp.py on a mock recording and run it with roboloom infer, and see where your own model goes.
  • Prerequisites: Python 3.12–3.14, uv, and a clone of the roboloom repository (next to the project, ../roboloom). You have read Installation. No hardware.
  • Time: 30 minutes.
  • Result: the policy drives a mock Rakuda2 and one episode is saved in the evaluation dataset datasets/rakuda2-eval.

Use a separate trial project

Even with --mock, the output is the same .roboloom/rakuda2/datasets/ as on hardware. Do not try this in a project that holds real data; create a new one.

flowchart LR
  A["mock recording<br/>record --mock"] --> B["train<br/>chunk_mlp.py train"]
  B --> C["checkpoint"]
  C --> D["run<br/>infer --mock"]

1. Create a trial project

uv init --python '>=3.12,<3.15' my_policy_trial && cd my_policy_trial
uv add git+https://github.com/tetsugo02/roboloom.git --tag v0.2.0 --extra rakuda2 --extra record --extra policy
uv run roboloom init rakuda2
cp ../roboloom/examples/policies/chunk_mlp.py .roboloom/rakuda2/   # from a clone of the repository
  • chunk_mlp.py is in the repository's examples/policies/, not in the installed package. Copy it from a clone of the same version as the roboloom you installed. A policy module in the robot directory (.roboloom/rakuda2/) is found without PYTHONPATH.
  • The policy extra brings torch and LeRobot.

2. Record with mock devices

uv run roboloom record --mock --task demo --seconds 20 --episodes 3 --plain
uv run roboloom check-dataset --mock

Check episodes: 3, ... fps: 15 and problems: none. Without --mock, loading is refused because SET_... values are left.

3. Train

uv run python .roboloom/rakuda2/chunk_mlp.py train \
  --dataset .roboloom/rakuda2/datasets/rakuda2 --out outputs/chunk_mlp.pt \
  --history 4 --horizon 8 --stride 3 --exclude torso_yaw head_yaw head_pitch
Argument Meaning
--history 4 the last 4 frames (one per policy period) are the input
--horizon 8 8 steps come out at once
--stride 3 train on every third row; on a 15 fps recording the policy runs at 5 Hz
--exclude ... joints the model does not drive; roboloom holds them (hold)

It prints saved outputs/chunk_mlp.pt: 14 of 17 joints, 5 Hz, after a few seconds on a CPU.

4. Run it on mock devices

Set rate_hz in .roboloom/rakuda2/nodes/policy.yaml to 5 (to match --stride 3). The history length, the held joints and the steps per chunk come from the policy's options().

uv run roboloom infer --mock --seconds 10 --plain --task "chunk eval" \
  --policy chunk_mlp:load --policy-kwargs "{\"checkpoint\": \"$PWD/outputs/chunk_mlp.pt\"}"
  • saved episode 0: 'chunk eval', ... frames means it went from recording to inference.
  • Pass the checkpoint as an absolute path. A relative path is not resolved from the project root (the dataflow is started from a temporary directory).
  • The mock recording hardly moves, so this says nothing about the policy's quality. It checks the path only.

5. Replace it with your model

Make the object load returns in chunk_mlp.py your own model. It needs only reset() and infer(observation).

class MyPolicy:
    def __init__(self, checkpoint: str): ...
    def reset(self): ...
    def infer(self, observation):
        return self.model(observation)  # [n] or [H, n]: absolute joint targets


def load(checkpoint: str) -> MyPolicy:
    return MyPolicy(checkpoint)
Step Where it goes
normalization, image preprocessing, turning the model's action form back into targets the policy
observation history, held joints, interpolation, speed limits, safety checks roboloom

Details are in Your own policy.

Next