Skip to content

Velocity task improvements#833

Open
kevinzakka wants to merge 10 commits intomainfrom
velocity-mdp-tweaks
Open

Velocity task improvements#833
kevinzakka wants to merge 10 commits intomainfrom
velocity-mdp-tweaks

Conversation

@kevinzakka
Copy link
Collaborator

@kevinzakka kevinzakka commented Mar 20, 2026

WIP

The encoder_bias DR event was applied but the actor observed true joint positions, making the bias invisible to the policy during training.
peak_heights was never cleared on episode reset, so stale values from the previous episode carried over and corrupted the metric.
@kevinzakka kevinzakka force-pushed the velocity-mdp-tweaks branch from 8f702e1 to e3cc8d0 Compare March 20, 2026 21:16
@kevinzakka kevinzakka force-pushed the velocity-mdp-tweaks branch from e3cc8d0 to fe093a8 Compare March 20, 2026 22:00
Comment on lines +29 to +30
cfg.sim.njmax = 200
cfg.sim.nconmax = 30
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wow, such a low njmax/nconmax work?

weight=-0.1,
params={"sensor_name": shank_ground_cfg.name},
)
cfg.rewards["thigh_collision"] = RewardTermCfg(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

WDYT about an alive reward? I was thinking it could be helpful especially for rough terrain

When a foot penetrates slightly below the terrain surface, all
downward rays miss and the sensor reported max_distance instead of ~0.
This corrupted foot clearance rewards and caused heel-walking gaits.

Use frame_z clamped to [0, max_distance] as the fallback for
all-miss frames instead of unconditionally returning max_distance.
Create Mjlab-Velocity-Flat-Run-Unitree-G1 task that extends the flat
config with a velocity command curriculum for learning to run. Move
the curriculum out of the base velocity config where it was applied
to all robots. Set Go1 flat iterations to 1.5k.
Increase base ang_vel_z from (-0.5, 0.5) to (-1.0, 1.0). Ramp the
G1 run curriculum stages to (-1.5, 1.5) and (-2.0, 2.0).
@kevinzakka kevinzakka force-pushed the velocity-mdp-tweaks branch from 2dc22bb to ab4286f Compare March 21, 2026 06:09
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants