Microduck — Raised-platform backflip

⚠️ Landings may break the robot. Simulation only; not yet validated on hardware. Successful simulated recovery does not establish a safe physical landing.

A backflip from a raised platform onto a soft-contact simulated mat. This is not a standing-height backflip and has not been validated on hardware.

Policy and landing assumptions

policy.onnx and checkpoint.pt: retained f21, iteration 12,000.
61 observations, 14 bounded joint-position actions, 50 Hz; normalization and action bounds included in ONNX, no action low-pass filter.
Body-command block: [(drop_height - 0.7) / 0.2, direction_sign, 0, 0, 0, 0]. The profile sets the backflip direction. Full contract: config.json.

The release profile uses a 0.8 m drop from platform top to mat top. The mat is 0.44342 m thick, with a 1.90154 m × 0.605 m footprint. Its collision is a soft-contact box (solref=(0.03, 1), solimp=(0.9, 0.95, 0.01, 0.5, 2)), not a calibrated model of a real crash mat. Servo and structural damage are not simulated. Do not extrapolate this result to a rigid-floor landing.

Preparatory bounces are part of the behavior. The reproduction script starts on the platform, not mid-air, and does not switch to a separate get-up policy. The sculpted visual mat in the preview is optional decoration; reproduction uses the same collision box as the visible mat.

Reproduce in simulation

Source and instructions

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

The preview is an excerpt from the selected research montage. Its exact checkpoint/seed mapping has not been verified; it is not an exact replay or success-rate claim for this package. Earlier research success figures affected by mid-air resets are not used as validation. See eval/ for numerical ONNX parity and a small fresh diagnostic of this checkpoint, manifest.json for provenance and SHA256SUMS for integrity. Load the pickle checkpoint only if you trust its source.

Related: long jump.

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

Completes the backward rotation and recovers to standing on the simulated mat, but later drifts off and falls at 7.62 seconds. This is not a stable idle policy. 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.

Downloads last month: 27

Video Preview

Reinforcement Learning