Dropbear whole-body control: plant, trained policies, media and motions
Assets for Hyperspawn/dropbear-wbc, the whole-body-control and
sim-to-real stack of Dropbear, an open 3D-printed humanoid robot. The code, documentation and install guide live in
the GitHub repository; this repository holds the files too large for git. Fetch them with the repository's
python tools/fetch_assets.py (SHA-256 verified against assets_manifest.json).
Try it without installing anything: browser replays of the motor twin.
| folder | contents |
|---|---|
robot/dropbear.usd |
the plant (421 MB USDC, SHA-256 45586414...): 90 bodies, 22 motors, 27 loop-closure joints |
policies/<run>/ |
trained rsl-rl checkpoints + run_info.json (actuator profile, calibration, target clamp) + params/*.yaml |
media/ |
dashboard videos: physics rollout + per-motor torque against the motor datasheets |
motions/ |
redistributable retargeted clips (Kimodo-generated, Kimodo G1 examples, ASAP), NPZ + validation verdicts |
datasets/groot/ |
GR00T tabletop push dataset (LeRobot format, simulated) |
All results are simulation results on a motor twin: Isaac Lab 2.2 / Isaac Sim 5.0, datasheet torque-speed envelopes, friction, latency, the 20 ms target ramp the firmware is meant to implement, and the robot's 27 closed kinematic loops. They are not measurements on the real robot.
Policies
| run / checkpoint | task | trained on | actuator profile | verified result (sim) |
|---|---|---|---|---|
lib_v7gen_v1ie_smooth/model_13600 |
tracking library, no-state obs | accepted_v7gen: 138 clips (60 Kimodo-generated, 51 ASAP, 10 BONES-SEED, 6 synthetic, 5 Kimodo G1, 4 LAFAN1, 2 unitree_rl_lab) |
hw_v1ie |
94/138 clips with 0 falls (51/60 generated, 43/78 original); default tracker of the live text-to-motion session |
lib_v6ts_v1i_fix2/model_13700 |
tracking library, no-state obs | accepted_v6ts: 78 clips (51 ASAP, 10 BONES-SEED, 6 synthetic, 5 Kimodo G1, 4 LAFAN1, 2 unitree_rl_lab) |
hw_v1i |
46/78 clips with 0 falls |
lib_v6ts_nostate_v1i_fix/model_10492 |
tracking library, no-state obs | accepted_v6ts |
hw_v1i |
41/78; recorded live session: 0 falls, HW gate PASS |
kimodo_ts119_allfix_smooth/model_5800 |
single-clip tracking, no-state | Kimodo G1 walk (Froude-timed cyclic); warm-started from earlier Kimodo-walk trackers | hw_v1ie |
walk: HW gate PASS, 0 falls, body error 6.2 cm |
lafan_nostate_smooth_elbow/model_8000 |
single-clip tracking, no-state | LAFAN1 walk1 (cyclic) | hw_v1ie |
HW gate PASS, 0 falls; right elbow 1.25x rated |
lafan_cyclic_nostate_v1i_kstop_feet/model_3100 |
single-clip tracking, no-state | LAFAN1 walk1 (cyclic) | hw_v1i (CEM-60 knee) |
HW gate PASS: knee stop contact 0.2 %, knees 0.94x / 0.82x rated |
vel_rough_v1i_b/model_5500 |
velocity, rough terrain (blind) | no motion data (reinforcement learning only) | hw_v1i |
HW gate PASS, 0 falls; limps, knees near the extension stop |
vel_hw_v1_gait3b_ft/model_1300 |
velocity, flat | no motion data | hw_v1 |
flat-ground gait-shaped walker |
hw_v1 = datasheet motor map; hw_v1i = + 4-step (20 ms) target ramp; hw_v1ie = + stiffer elbow gains. "HW gate" =
tools/telemetry_issue_scan.py (fails on falls, thermal overload, clipping, over-speed, hard-stop loading, skids,
slips, scuffs).
Licenses and provenance
- This repository (plant USD, policies, videos, datasets) and the code that produced it: noncommercial use only,
PolyForm Noncommercial License 1.0.0 (
LICENSE). For commercial use, contact Hyperspawn: hyperspawn.co, priyanshu@hyperspawn.co. - Motions here: retargeted from Kimodo-generated / Kimodo G1 motions (outputs of NVIDIA Kimodo-G1 under the NVIDIA Open Model License) and ASAP motions (MIT); the retargeted clips are this project's (noncommercial). Clips from LAFAN1 (CC BY-NC-ND 4.0), NVIDIA BONES-SEED sample data (evaluation license) and unitree_rl_lab dance mocap are not included; the repository documents how to rebuild them from your own download.
- Policies were trained on the libraries listed above, which include LAFAN1 and BONES-SEED sample clips, and two
are LAFAN1-walk trackers. They are released for noncommercial research use only. For a clean-lineage library
policy, train on
accepted_v7gen_public.json(redistributable clips only). - Videos show the simulated robot performing those motions; the same provenance applies.
Use
git clone https://github.com/Hyperspawn/dropbear-wbc && cd dropbear-wbc
pip install huggingface_hub && python tools/fetch_assets.py # USD + policies
python tools/fetch_assets.py --groups media,motions,datasets # everything else
Then follow docs/INSTALL.md and docs/QUICKSTART.md in the repository.