Skip to content

One RoboCasa sim loop for both robots (golem_sim.py); Isaac elastic band - #37

Merged
maxconwa merged 2 commits into
mainfrom
golem-sim-unify
Nov 20, 2026
Merged

maxconwa merged 2 commits into
mainfrom
golem-sim-unify

Conversation

@maxconwa

Copy link
Copy Markdown
Contributor

What this does

One RoboCasa sim loop. golem_robocasa/golem_sim.py is now the entry point and the only physics loop for both robots. What differs per robot is one object each with the same interface: h12_robot.H12Robot (was h12_mujoco.py) and g1/g1_robot.G1Robot (was g1/g1_mujoco.py). A change to the loop is made once.

  • Each robot keeps the locking and camera behaviour it had.
  • Both robots now log the real-time factor every 30 s of sim time.
  • --timing and --duration are removed, with the per-stage timing report.
  • The G1's spawn search now also checks its limp arm poses and its start-up arm move, as the H1-2's does. The G1's limp poses were measured under the band.

Isaac elastic band. --elastic_band hangs the floating H1-2 from the same band as the RoboCasa H1-2 sim and serves /elastic_band/toggle. It is off by default and refused with --fix_base. This bumps CL_isaaclab_sim to its elastic-band branch (78c3f70); merge that branch there first, or re-pin after it merges.

Tested (i9-14900KF, RTX 4090)

Check Result
Surface probes, H1-2 and G1 match the goldens
tests/sim/test_sim_stack.sh --h12 / --g1 14/14 each, 1.00x
tests/g1: arms, walk, nav 22/22, 22/22, 13/13
tests/stack/run.sh robocasa arms / locomotion pass
RoboCasa unit tests 200 H1-2, all 7 G1 files
tests/docker, tests/stack host suites pass
Isaac unit tests; tests/isaaclab/run_elastic_band.sh pass; 4/4 (40 s hold, release, re-hold)
Unpaced speed H1-2 about 2.4x, G1 about 3.7x, same as before

Failing as on main before this change:

  • tests/g1/test_sim_explore.sh: 9/10, coverage (the open issue in tests/g1/README.md).
  • tests/sim/test_lmg_explore.sh: 3/8, the robot fell 2 s after the band's release; it fell in 2 of 4 runs on main today.

Not tested

  • The Mac. It runs the same golem_sim.py through the same launcher, and the CPU-caster and wrapper unit tests pass, but nothing was run on a Mac.
  • The ROS package tests were not rerun; core_ws is untouched.

🤖 Generated with Claude Code

https://claude.ai/code/session_015EG7p9fHRyH9uB5cMqdm8P

maxconwa and others added 2 commits November 19, 2026 20:13
…check

Bumps CL_isaaclab_sim to its elastic-band branch (78c3f70): --elastic_band
hangs the floating H1-2 from the RoboCasa H1-2 sim's band on torso_link until
/elastic_band/toggle releases it. Off by default, refused with --fix_base.

tests/isaaclab/test_elastic_band.py checks the force, the toggle and what is
stored on the robot without Kit. tests/isaaclab/run_elastic_band.sh runs the
real sim for 50 s with no controller and calls the service over ROS: the limp
robot hangs within 0.10 m of the anchor for 40 s, collapses on release, and
is pulled back from where it lay when the band is enabled again.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015EG7p9fHRyH9uB5cMqdm8P
…ct each

golem_robocasa/golem_sim.py is the entry point and the one physics loop.
What differs per robot is an object with one interface, h12_robot.H12Robot
(was h12_mujoco.py) and g1/g1_robot.G1Robot (was g1/g1_mujoco.py): spawn,
band, sensors, DDS interface, locking, and the H1-2's grippers and estop box.
Each robot keeps the locking and camera behaviour it had. Both now log the
real-time factor every 30 s of sim time and handle MuJoCo's instability
reset (LowCmdLatch.physics_reset).

--timing and --duration are gone, with the H1-2's per-stage timing report
(LoopTimings) and its plumbing; sim_loop keeps a duration argument for the
unit tests.

The G1's spawn search now also keeps its margin in the limp poses its arms
pass through under the band (measured: the same in the Blank task and a
kitchen) and along the controller's start-up arm move, as the H1-2's does
(robosuite_common.startup_arm_poses; tests/docker/test_sim_startup_poses.py
covers both robots).

CLAUDE.md, the G1 docs and the test READMEs name the new files, and CLAUDE.md
describes the Isaac band.

On an i9-14900KF with an RTX 4090: both surface goldens match; tests/sim
14/14 for both robots at 1.00x; tests/g1 arms 22/22, walk 22/22, nav 13/13;
tests/stack arms and locomotion pass; 200 H1-2 and all G1 sim unit tests
pass. Unpaced, the H1-2 runs at about 2.4x and the G1 at about 3.7x, as
before. Not run on the Mac.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015EG7p9fHRyH9uB5cMqdm8P
@maxconwa
maxconwa merged commit 4656a07 into main Nov 20, 2026
4 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