Downloads · 30 days
0
hvsr-robotics/wojtek-quiet-locomotion
wojtek-quiet-locomotion is a machine learning model from hvsr-robotics. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Quiet-walking keeper for Wojtek: wojtekflatquietv10s0 (2026-08-12). Soft-landing gait on every linear command, step-turn spins (no jump-turns), trained without any gait choreography — the quietness comes from priced p…
Downloads · 30 days
0
Access
Public
Updated Aug 25, 2026
Repo size
4.5 MB
Likes
1
Public
Click a slice to open those files.
.npz2.3 MB · 51%
From the Hugging Face model README
Quiet-walking keeper for Wojtek: wojtek_flat_quiet_v10_s0
(2026-08-12). Soft-landing gait on every linear command, step-turn spins
(no jump-turns), trained without any gait choreography — the quietness
comes from priced physics, not a clock.
| metric | sag_v3_s1b | v10 |
|---|---|---|
| touchdown speed p90, walk 0.5 (m/s) | 0.89 | 0.23 |
| free-fall ratio, slow/walk/fast | 0.88 / 0.90 / 1.14 | 0.09 / 0.24 / 0.29 |
| spin gait | jump-turn, 17.7 taps/s, 14% airborne | step-turn, 6.5 taps/s, 0% airborne |
| spin touchdown p90 (m/s) | 1.13 / 0.94 | 0.28 / 0.34 |
| achieved wz (spin L/R) | 0.96 / −0.98 | 0.90 / −0.88 |
| courses falls | 1–2 / 160 | 0 / 160 (slippery circle 8/8 — first in lineage) |
| actuator saturation (battery) | — | 0 everywhere |
Recipe (all committed on michal/flat-tff-rnd, presets
flat_quiet_v1..v6 + CLI overrides recorded in the report):
keeper-curriculum restore + feet_landing=-15 with
feet_landing_wz_fade=0.1 / feet_landing_wz_floor=0.6 (spins pay 60%)
energy_cmd=1.0 (arXiv 2403.20001 command-normalized energy reward,
tracking-gated) + tracking_ang_vel=2.4 (defends spin rate). Seed-1
replication lands within noise. Full experiment log:
QUIET-REPORT-2026-08-12.md.policy.npz + policy_meta.json — the schema-2 deployment pair
(24-wide action: position targets + tau_ff head). Deploy by reference:
policy:=hvsr-robotics/wojtek-quiet-locomotion on the launch files.checkpoint/ — the final orbax checkpoint (001002700800), restorable
with restore= for further fine-tuning.battery.json / eval_report.md — the full battery report.courses.json + courses/*.png — 20-scenario path-following
benchmark (8 seeds) and overhead trajectory plots.variants/v9 — floor 0.475 arm (spin td 0.28/0.31 at wz 0.84/−0.87).variants/v2b — the quietest linear gait of the family (ffr
0.15–0.32 everywhere including gentle arcs) but pivoting eroded to
0.05 rad/s; only for use-cases that never turn in place.Lineage: flat_tff_rnd_v1 (tau_ff head + RND, kp20/kd0.5) → flat_tff_sag_v2b (gravity-compensation sag reward) → flat_tff_sag_v3 (real foot-friction DR) → quiet family (this repo). Trained with MuJoCo MJX/MJWarp + brax PPO, 2×RTX 5090, ~580k SPS; base run 2B steps, quiet fine-tunes 1B.