Wheel Leg Robotic / CJ-003

A two-wheel biped that is actually clonable on a Mac.

We searched public wheel-legged stacks, skipped anything that requires NVIDIA Isaac, cloned wheel_legged_genesis, and wired its pretrained PPO policy into MuJoCo.

Albusgive/wheel_legged_genesis MIT policy 286 → 8 6 terrains MuJoCo 3.3 · CPU

What we looked at

Good demos are easy to find. Easy reproduction on Apple Silicon is not. Isaac Gym / Isaac Lab were treated as a hard no.

Project	Demo	Why not / why yes
DreamWaQ Go2-W / LocoLeggedWheel	Strong real-robot Go2-W video	NVIDIA Isaac Gym or Isaac Lab.
SUSTech Wheel-Legged-Gym	Classic wheeled biped RL	Isaac Gym Preview 4.
jaykorea Flamingo (Isaac + Genesis forks)	Excellent sim2real	Isaac Lab; Genesis fork still heavy.
LimX TRON WF MuJoCo	Vendor wheel-foot humanoid	Linux SDK wheels only.
Albusgive/wheel_legged_genesis	Bilibili locomotion + MuJoCo transfer	MIT, in-repo checkpoints, MuJoCo XML, no NVIDIA required for replay.

The robot

CJ-003 two-wheel biped in MuJoCo and terrain — Upstream sim2sim stills: checkered plane and rough terrain. Two hip–thigh–calf chains, driven wheels, four-bar linkages.

Genesis training views of the wheel-legged robot — Genesis training views from the original repository, including bumpy evaluation terrain.

CJ-003 is a balancing wheel-legged biped: six position-controlled leg joints and two velocity-controlled wheels. Commands are forward speed, yaw rate, left/right leg length, and a lean term.

The v0.3.2 policy is a TorchScript actor, 286 observations (9 history frames + current 28-D slice + 6 scaled commands) to 8 actions.

nq 19 nv 18 nu 8

slice 28 = ω(3)+g(3)+q_leg(6)+dq(8)+a(8)

dt 0.002 physics · 5× decimation

ckpt logs/wheel-legged-walking-v0.3.2/policy.pt

Motion repertoire

Everything this v0.3.2 actor can track through its 6-D command: vx, vy, yaw, left_leg, right_leg, lean. vy was trained at 0, so there is no lateral strafe. No jump channel either.

CUDA_VISIBLE_DEVICES=1 /data/ypx/.conda/envs/wl_genesis/bin/python \
  reproduce/run_genesis_motion_demo.py
# full gallery: media/motions/

Trained checkpoint on different terrains

Genesis GPU eval of the official v0.3.2 actor (policy.pt from model_21000.pt). This is the training simulator on CUDA, not MuJoCo. Scripted forward command 0.8 m/s. All three maps stayed upright in this run.

Genesis eval gym — Eval gym · stayed up · 3.37 m · min z 0.67

Genesis train gym — Train gym · stayed up · 3.67 m · min z 0.66

Genesis circular slopes — Circular · stayed up · 3.93 m · min z 0.67

CUDA_VISIBLE_DEVICES=1 /data/ypx/.conda/envs/wl_genesis/bin/python \
  reproduce/run_genesis_terrain_demo.py --seconds 8
# MuJoCo heightfield replay still falls (sim2sim gap):
python reproduce/run_terrain_demo.py

What we reproduced here

Cloned the repo, loaded the shipped policy on CPU PyTorch, patched the MJCF (Linux absolute paths → relative; planar slides → freejoint), and stepped the closed-loop controller in MuJoCo 3.3 from the existing twip_rl conda env.

The loop runs. Visual offscreen recording is blocked in this agent session (no CoreGraphics), and mesh-contact sim2sim still dumps the robot after a few hundred ms — the upstream Python sim2sim file is still the older 6-action observation. The Mac replay script uses the v0.3.2 8-action layout.

On a local Mac display you can open the MuJoCo viewer from reproduce/run_mujoco_demo.py. Official GPU training/eval stays on Linux + Genesis, as the authors specify.

Run it

# motion repertoire (Genesis + CUDA):
CUDA_VISIBLE_DEVICES=1 /data/ypx/.conda/envs/wl_genesis/bin/python \
  reproduce/run_genesis_motion_demo.py
# terrains:
CUDA_VISIBLE_DEVICES=1 /data/ypx/.conda/envs/wl_genesis/bin/python \
  reproduce/run_genesis_terrain_demo.py --seconds 8
# MuJoCo CPU replay (will tip — known sim2sim gap):
python reproduce/run_mujoco_demo.py --seconds 8

Keyboard command map — Upstream keyboard map for Genesis eval.

Gamepad command map — Gamepad map: left stick speed, right stick yaw, triggers for height.

Homepage · ypx19 · Source: Albusgive/wheel_legged_genesis · MIT · local clone in wheel_legged_genesis/ · page generated for the Wheel Leg Robotic workspace.