Success rate of climbing tasks

[baijubaiju233]

In the Go2 crate climbing task, I observed that the success rate is much lower than the 0.9 mentioned in the paper. To address this, I attempted to adjust the height of the crate and modified the reward parameters as follows:

Humanoid Crate Pushing Reward Table

Gait: 5.0

Upright: 0.01

Yaw: 0.1

Velocity: 1.0

Torso Height: 0.5

Energy: 0.01

Contact Reward: 0.05

However, the Go2 robot only succeeded 1 or 2 times out of 10 attempts with an unstable posture. How can I further revise the code to improve the success rate of this climbing task?

[HaoruXue]

@baijubaiju233 thanks for the discovery - some recent package updates might have affect the results.

without changing the rewards, can you try playing with the parameters in dial_mpc/examples/unitree_go2_crate_climb.yaml? Some settings that might be helpful to tune are:

Hsample: horizon length
Hnode: number of spline node. must be an integer factor of Hsample
Ndiffuse: number of diffusion steps
traj_diffuse_factor: variance schedule at every diffusion step

For example, increasing Ndiffuse to 3 and traj_diffuse_factor to 0.9 yields resonable results on my side.

Feel free to play around and find a good number to put in a PR! Thank you.