Skip to content

lmg_deploy: policies/h12 is LegsEnergyDefault200k_s43; a legs actor without the wrist targets - #39

Merged
maxconwa merged 1 commit into
mainfrom
lmg_deploy/legs-energy-policy
Nov 20, 2026
Merged

maxconwa merged 1 commit into
mainfrom
lmg_deploy/legs-energy-policy

Conversation

@maxconwa

Copy link
Copy Markdown
Contributor

What changes

  • policies/h12 holds a new policy, and only that one. LegsEnergyDefault200k_s43 (checkpoint agent_200160, task LocoManip-LegsOnly-v0) replaces LegsBalanceAmp_s42; the old networks and export are gone from the folder. It is the same sport-mode recipe at 100 Hz, legs only, trained with a mechanical-power penalty and no pelvis-height reward.
  • The controller accepts a legs actor that is not shown the wrist targets. Locomanipulation_game's legs-only runs since 2026-11-16 drop ee_targets from the legs' observation (83 inputs instead of 97). controller.py now accepts a legs_obs_terms list without ee_targets and builds the legs' observation without them. The arms' list must still be the full one.
  • Docs: README.md names the new policy and the shorter legs layout; config/example.export.yaml documents three keys newer exports write (residual_cutoff_hz, ik_damping, clip_actions).
  • h12.yaml: legs.num_obs is 83.

How it was checked

Check Result
lmg_deploy package tests (ros container) 54 passed, 1 skipped
Locomanipulation_game deploy_parity.py: this controller against the training env in Isaac, 600 steps passes; worst difference 5.1e-5 (arm targets)
tests/sim/test_lmg_explore.sh (RoboCasa, FAST-LIO, Nav2) 8 of 9

The RoboCasa run: no fall, no contact with anything but the floor, no e-stop during start-up, tilt at most 9.5 deg, both Nav2 goals reached. The failing check is coverage: the explorer reported the map explored after 2 goals with 41.6% of the floor mapped (limit 95%). Two earlier runs of the same test that day with the previous policy also failed coverage, with the explorer stopping early, so this looks like the explorer rather than the policy; that is not confirmed.

Known limits of this policy

  • It walks in a crouch (pelvis height bottomed out at 0.770 m in the RoboCasa run, the test's lower limit).
  • Scorecard at 200k steps (84 envs): turn sense 0.976, direction cosine 0.973, linear error 0.053 m/s, angular error 0.107 rad/s, no falls. Turning is weaker than sport mode's (0.994, 0.050 rad/s).
  • Not run on the robot.

🤖 Generated with Claude Code

https://claude.ai/code/session_015EG7pDM5uvsCpuEnXVWvMd

…ithout the wrist targets

policies/h12 now holds LegsEnergyDefault200k_s43 (agent_200160, LocoManip-LegsOnly-v0),
replacing LegsBalanceAmp_s42: the same sport-mode recipe, trained with a
mechanical-power penalty and no pelvis-height reward.

Its legs actor is not shown the wrist targets (83 inputs, not 97), so the
controller accepts a legs_obs_terms list without ee_targets and builds the legs'
observation without them; the arms' list is unchanged. example.export.yaml
documents the three keys newer exports write (residual_cutoff_hz, ik_damping,
clip_actions).

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015EG7pDM5uvsCpuEnXVWvMd
@maxconwa
maxconwa merged commit f7f1bda into main Nov 20, 2026
2 checks passed
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.

1 participant