Unitree Go2 Modified Calf Gait RL Simulation on DUVEE model
Budget: $30 – $250 USD
I have working MuJoCo models of a customized Unitree Go2, where the CALF geometry has been structurally changed, stretched, and re-balanced. The walking and trot gaits are based on the original Go2, but the CALF structure has been modified, including changes in bending direction and joint angles.
In particular, the front-leg CALF has a completely inverted structure compared to the original Go2, bending inward instead of outward.
In addition, basic sit-to-stand and stand-to-sit motions must be implemented as a prerequisite, since stable transitions into and out of locomotion are essential.
Although walking and trot behaviors already run, the robot still wobbles, tips over, or drifts after a few meters. My priority is to push these gaits through a full reinforcement-learning training cycle until they remain rock-steady on uneven virtual terrain.
Where things stand
• Joint limits and contact parameters are roughly in place, yet the trade-off between stability and freedom is still delicate—too loose and the legs splay, too tight and the RL agent plateaus.
• Dynamics have been validated against CAD mass properties, but I’m open to a second pair of eyes if you spot inconsistencies.
• Training is done with MuJoCo + Python (mujoco-py), scene_terrain.xml
What I need from you
1. Review and refine the joint-limit definitions so the agent can safely explore without sacrificing natural stride length.
2. Design or tweak a reward function that explicitly favours stability over long horizons for walking and trot.
3. Run or script RL experiments (preferably on GPU) until each gait reaches consistent forward velocity ≥ 0.8 m/s with less than 3 cm lateral drift over 10 m.
4. Hand back clean, commented code plus a short report (metrics, screenshots or brief demo video).
Tools you are comfortable with—MuJoCo, RLlib, Stable-Baselines, Isaac Gym, JAX—are all welcome as long as the final solution plugs into my current Python workflow.
Acceptance check
• Both gaits complete 50-m virtual runs on flat and mild 15° slopes without falls.
• Re-running training from scratch with the supplied scripts reproduces ±5 % of your reported reward curve.
• Joint torques stay within manufacturer specs.
Up to a speed of 1.5m/s, the robot uses a walk gait.
When the speed is in the range of 1.5–1.6 m/s, it transitions into an intermediate trot gait with a smooth takeoff.
At speeds higher than that, the robot switches to a full trot gait.
The stop, sit, and transition from standstill to walking are designed so that all gaits are smoothly and continuously connected without abrupt changes.
If this sounds like your kind of challenge, let’s stabilise this robot together.
https://github.com/felixokolo/go2_gait_planner
https://github.com/unitreerobotics/unitree_mujoco/tree/main/unitree_robots/go2
https://www.youtube.com/shorts/bbROvIr_dDw trot animation
https://www.youtube.com/shorts/I--IpZ2Wjxs walk animation
In particular, the front-leg CALF has a completely inverted structure compared to the original Go2, bending inward instead of outward.
In addition, basic sit-to-stand and stand-to-sit motions must be implemented as a prerequisite, since stable transitions into and out of locomotion are essential.
Although walking and trot behaviors already run, the robot still wobbles, tips over, or drifts after a few meters. My priority is to push these gaits through a full reinforcement-learning training cycle until they remain rock-steady on uneven virtual terrain.
Where things stand
• Joint limits and contact parameters are roughly in place, yet the trade-off between stability and freedom is still delicate—too loose and the legs splay, too tight and the RL agent plateaus.
• Dynamics have been validated against CAD mass properties, but I’m open to a second pair of eyes if you spot inconsistencies.
• Training is done with MuJoCo + Python (mujoco-py), scene_terrain.xml
What I need from you
1. Review and refine the joint-limit definitions so the agent can safely explore without sacrificing natural stride length.
2. Design or tweak a reward function that explicitly favours stability over long horizons for walking and trot.
3. Run or script RL experiments (preferably on GPU) until each gait reaches consistent forward velocity ≥ 0.8 m/s with less than 3 cm lateral drift over 10 m.
4. Hand back clean, commented code plus a short report (metrics, screenshots or brief demo video).
Tools you are comfortable with—MuJoCo, RLlib, Stable-Baselines, Isaac Gym, JAX—are all welcome as long as the final solution plugs into my current Python workflow.
Acceptance check
• Both gaits complete 50-m virtual runs on flat and mild 15° slopes without falls.
• Re-running training from scratch with the supplied scripts reproduces ±5 % of your reported reward curve.
• Joint torques stay within manufacturer specs.
Up to a speed of 1.5m/s, the robot uses a walk gait.
When the speed is in the range of 1.5–1.6 m/s, it transitions into an intermediate trot gait with a smooth takeoff.
At speeds higher than that, the robot switches to a full trot gait.
The stop, sit, and transition from standstill to walking are designed so that all gaits are smoothly and continuously connected without abrupt changes.
If this sounds like your kind of challenge, let’s stabilise this robot together.
https://github.com/felixokolo/go2_gait_planner
https://github.com/unitreerobotics/unitree_mujoco/tree/main/unitree_robots/go2
https://www.youtube.com/shorts/bbROvIr_dDw trot animation
https://www.youtube.com/shorts/I--IpZ2Wjxs walk animation
Related categories:
Python
3D Modelling
Robotics
Simulation
Robot Operating System (ROS)
Reinforcement Learning