Conversation
Adds the g1-wbt-backflip WBT task: Unitree G1 tracks a full backflip reference clip (converted from the mjbatch flip motion with FK cross-check) and learns the complete skill — takeoff, full rotation, landing, and stable hold — under motrix.fastsac. Validated recipe (each element from controlled comparisons): - mixed reset sampling: 20% frame-0 starts + 80% uniform whole-clip RSI (teleports exact in velocity: root velocity reset noise zeroed for this task; centered reset noise everywhere else). Frame-0-only start starves the flight window of episode-start coverage; adaptive sampling is strictly harmful in BOTH regimes — the uniform_ratio sweep (0.1/0.5/0.8) lost to plain uniform from scratch at every value (40k rotation 0.08/ 0.15/0.06 rad vs 0.48) and fine-tuning the flipping 80k policy with adaptive sampling regressed the play-visible skill - sim dt 0.005 required: dt=0.01 doubles collector throughput but the ~0.1 s launch impulse is lost to the coarse contact solve and rotation never forms - motion_global_body_ang_vel at the UniLab sigma (3.14); earlier tightening to 1.0 saturated the kernel to ~0.02 across whole runs - flight-window rewards projected onto the clip's flight axis: linear flight_rotation_progress (unsaturating, pays per radian) plus FlightTuckRewardCfg over all eight leg joints including hip roll/yaw abduction (sigma 1.0, weight 4.0 — open-leg flight inflated the moment of inertia and stalled rotation at ~3 rad; extended coverage reached a 3.8 rad peak at 60k, 2.2x faster than the knee/hip-pitch-only variant) and a full-3D EE term (weight 3.0, replacing the z-only variant: ablating z-only alone stalls launch bootstrap ~20k iters because the 3D error is dominated by early xy drift — the raised 3D weight keeps the height emphasis); action_rate_l2 flat -0.01 (sweep: every heavier flat or gated variant failed to unlock rotation) - phased terminations: anchor z loose outside the flight window / tight inside (0.5/0.25, closes the tiptoe exploit at the apex), phased 3D anchor position (0.8/1.0); holosoma-contract strict terms available as metrics-only thresholds for later fine-tuning Clip-data and metric fixes (the breakthrough changes): - g1_flip_converter: MotrixSim free-joint qvel takes world-frame LINEAR but BODY-frame ANGULAR velocity (MuJoCo convention); feeding world-frame omega corrupted every baked body_ang_vel_w as R @ omega while positions stayed correct (positions were the only cross-check). Fixed, plus a new FK velocity self-check guards the convention; backflip.npz re-baked (flight-window x-axis integral now ~ -2pi) - flight_axis: the flip rotation axis is derived from the reference window angular-velocity integral (signed, world frame); flight_pitch_rotation and flight_rotation_progress project onto it instead of a hardcoded world-y index — the robot faces -/+y in this clip, so world y was measuring the roll component. A perfect flip now reads ~2pi on the metric Honest-skill metrics in WbtMotionCommand (flight window derived from the clip root-z profile): flight_max_pelvis_z, flight_pitch_rotation (rotation about flight_axis), landed_upright. Known dilution caveats and the conditional-metric redesign are recorded in wiki/design/g1-backflip-flight-metrics.md. Also: pytest-xdist in the dev group (25 WBT tests: 5-8 min -> 2 s warm).
…blation plan A1 ablation (40k): rotation/height identical to phased baseline, landed_upright 0.65 vs 0.18. Details in wiki/plan/g1-backflip-ablation.md.
A3 ablation (40k): flight_tuck showed no measurable effect on rotation or height vs A1, play skill intact. Remove the term, its kernel, and the unreliable fixed-frame landed_upright metric (see wiki/design/g1-backflip-flight-metrics.md); record results in the ablation plan.
40k ablation showed no measurable effect vs zeroed noise; restore default holosoma reset noise and remove the task-specific __post_init__ hack.
Weight-0 leftover with no effect (A6a); delete the cfg, kernel, and stale comments. Record A6a outcome (term kept, landing accuracy regression) in the ablation plan.
…k, no play-level gain)
Episode-mean rotation is not interpretable under mixed/RSI sampling and is manipulated by start distribution (uniform vs adaptive gave the same policy a 3.7x different mean). Keep flight_max_pelvis_z and the flight_axis needed by the rotation-progress reward; see wiki/design/g1-backflip-flight-metrics.md for the binary success-rate replacement design.
…ation) Re-ablated on the current recipe (no flight_tuck): removal showed no regression in height, tracking error, episode length, or play behavior. Delete the term, kernel, and now-unused flight_axis plumbing. Document the flat ang-vel kernel finding (A9 candidate).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
g1-wbt-backflip) trained with Motrix FastSACscripts/private/g1_flip_converter.py), design notes (wiki/design/g1-backflip-flight-metrics.md), and dependency updatesTraining evidence
Play recording (best policy, 16 envs)
Reward vs walltime