Parameter recommendations for adapting from Go2 to Unitree A2?

[JiayingST]

Hi, we're reproducing the paper and adapting the pipeline to the Unitree A2
(40 kg, thigh/calf 0.275 m) instead of Go2. The code runs, but we're unsure
whether our parameter scaling is correct.

Our current changes vs. Go2 defaults:

Parameter	Go2	Our A2	Rationale
stiffness	40	120	Scaled by mass ratio (40/15 × 40 ≈ 107)
damping	1	3	Proportional to stiffness
init_state.pos[2]	0.42 m	0.45 m	A2 standing height
base_height_target	0.25 m	0.32 m	A2 base height
BBC jump_height	[0.45, 0.58]	[0.58, 0.75]	Scaled by leg length ratio (~1.29×)
BBC locomotion_height	[0.25, 0.34]	[0.32, 0.44]	Same scaling

Are there other parameters we should adjust (e.g., terrain obstacle sizes,
reward scales, action_scale, command velocity ranges)?

Also, is there a rough BBC quality threshold (episode_length / reward_i)
that should be reached before TSC training becomes stable?

BBC training result

Trained BBC for ~148k iterations with the above parameters:

• episode_length: ~378 (Go2 typically reaches ~1000)
• reward_i: ~2.71

Is this sufficient quality to start TSC training, or should we continue BBC longer?

TSC training: PhysX CUDA crash

With the BBC above, TSC teacher training crashes after ~3600 iterations with:
GPU compressContactStage1 fail to launch kernel stage 1!!
Gym cuda error: an illegal memory access was encountered

Our hypothesis: the BBC is still somewhat undertrained → TSC episode length is ~58 steps
→ high reset rate → some robot respawns into an obstacle → physics explosion.

Have you encountered this crash? Is there a known fix or PhysX config to avoid it?

Thank you!