Microduck — Raised-platform long jump

A learned jump from one raised platform to a lower platform across a gap. This uses the platform-jump task, not the flat-ground long-jump task.

Simulation only; not yet validated on hardware.

Policy

• policy.onnx: 61 observations → 14 bounded joint-position actions, 50 Hz.
• checkpoint.pt: retained p1 checkpoint, stored iteration 15,250.
• config.json: joint order, offsets, action bounds and command-slot meaning.
• eval/: numerical ONNX parity and a small fresh simulation diagnostic.

Observation normalization and action bounds are baked into ONNX; there is no action low-pass filter. The actor has no vision: the body-command block is [gap / 0.3, drop / 0.3, 0, 0, 0, 0], with distances in metres. Matching the walking policy's dimensions does not make it a drop-in replacement.

Reproduce in simulation

Source and instructions · Released jump experiments

From that source checkout, after downloading checkpoint.pt:

uv sync --locked
uv run python scripts/parkour_release.py long-jump render \
  --checkpoint /path/to/checkpoint.pt --output eval/long-jump.mp4

The supplied profile fixes a 30 cm gap and 25 cm drop, with the target platform 10 cm above the floor. It begins on the starting platform and stops at the first episode termination. No separate recovery policy is used.

The preview is a selected excerpt from the parkour montage. Its exact checkpoint/seed mapping has not been verified; it is not presented as a seeded replay of this package or evidence of a measured success rate. Evaluation files refer specifically to the packaged checkpoint. manifest.json pins the source revision, profile, checkpoint hash and media provenance; SHA256SUMS covers the package. Load the pickle checkpoint only if you trust its source.

Related: backflip.

Continue training

TRAINING.md documents the full-checkpoint continuation recipe, distinct from render profiles. A clean locked installation passed a 64-environment/five-update resume, exact learning-state restoration, finite-step checks and normalized ONNX parity. Recipe provenance and reconstruction limits are explicit; this is not a claim of bit-identical historical replay or improved behavior. Original policy, checkpoint and media are unchanged.

Fresh release check — 2026-09-29

Crosses the gap, lands and stays standing on the target through the 12-second rollout. One CPU rollout at the documented default seed (28), using the exact packaged checkpoint; not a success-rate estimate or hardware validation. See source verification and eval/render-20260929.json.