microduck-step-up-head-brake

Microduck Step-Up + Head-Brake Recovery

A community-trained two-policy controller for the Pollen Robotics Microduck: cross a 25 mm square-edged step, use the head as a temporary brake/support, recover upright, and stop on the upper platform.

Status: simulation-validated, hardware-unvalidated. This policy has not yet run on a physical Microduck. Use a support rig, tether, padded floor and emergency stop for first hardware trials.

Full source, training snapshot and documentation: bihaokun/microduck-step-up-policy

Preview

The full reference rollout is available in preview.mp4, with a still montage in preview_montage.png.

Policies

Both actors use the official Microduck runtime contract:

float32 obs[1,61] -> actions[1,14]
control rate: 50 Hz
observation normalization: baked into ONNX

This is a controller-triggered skill. It does not detect a step by itself. An operator or external perception layer must align the robot and set the velocity command to zero after both feet reach the upper surface.

Simulation evaluation

Test geometry: 25 mm high, 800 mm wide, square-edged step.

Stable means both feet remain on the upper platform, upright cosine is at least 0.8, and horizontal speed is at most 0.05 m/s at the end. The known failure starts 2 cm closer at -10 degrees yaw: the robot reaches the upper surface, then exits through the platform side.

These are deterministic simulation cases, not a statistical guarantee and not evidence of sim-to-real transfer.

Files and reproducibility

• step_up_walk_model_1000.pt: selected crossing PPO checkpoint.
• microduck_rl_step_up_source.zip: self-contained training snapshot with the Kaggle notebook, task source, reward functions, terrain, recovery-state bank, tests and bootstrap checkpoint.
• manifest.json: machine-readable policy contract and evaluation status.
• robotd-policy-overrides.toml: guarded hardware-deployment candidate.
• SHA256SUMS: integrity hashes for every uploaded artifact.

The training stack is based on pollen-robotics/microduck_rl at base revision d424a0c899f6b33cbd3daeb279913134349c0b63.

Known provenance limitation: the exact PyTorch checkpoint used to export the selected recovery ONNX was not retained. Its deployable ONNX, recovery task, state bank and retraining code are included, but that exact PPO run cannot be resumed. The crossing actor can be resumed.

Safety

Do not deploy this model to hardware without reading the repository's SAFETY.md. The policy can make a physical robot fall, collide, pinch, overheat or leave the platform. This is an independent community experiment, not an official Pollen Robotics release.