Skip to content

Repo cleanup - #9

Merged
maxconwa merged 179 commits into
mainfrom
repo_cleanup
Nov 10, 2026
Merged

maxconwa merged 179 commits into
mainfrom
repo_cleanup

Conversation

@maxconwa

Copy link
Copy Markdown
Contributor

Summary

  • One leg package for both robots: golem_lowerbody_rl replaces h12_lowerbody_rl and g1_homie. One config-driven node runs WALK, ALMI, FAME and HOMIE, checked against golden traces.
  • Sim speed: MuJoCo 3.17, a GPU-cast MID-360, and GOLEM_SIM_RATE (default 1×). Every golem node now runs on sim time under use_sim_time.
  • One platform: x86 Linux with an NVIDIA GPU. The Mac path, the ROS MCP server, MuJoCo MPC (deploy_mjpc, mujoco_mpc) and OSMesa are removed.
  • Cleanup: dead and duplicate code deleted, and comments rewritten to the CLAUDE.md rules.
  • CL_Assets: pinned to main's da59e8c, which uses the H1-2 MID-360's raw frame.
  • CI: Docs now builds on PRs and deploys only once GitHub Pages is enabled.

After pulling

  • sudo rm -rf core_ws/build/deploy_mjpc core_ws/install/deploy_mjpc (other stale installs: CLAUDE.md §10)
  • docker/scripts/docker_build.sh robocasa ros
  • The RoboCasa sims now need an NVIDIA GPU.

Tests

  • Clean build of all 21 packages; colcon test: 1230 tests
  • tests/docker 114, RoboCasa unit tests (H1-2 182, G1 7 files)
  • Full-stack sims: H1-2 14/14, G1 14/14
  • Stack arms (frame_task_server through golem_safety) and locomotion; G1 walk 22/22
  • Surface goldens match; Docs and CPU CI checks replayed locally
  • Watch both robots in the viewer on MuJoCo 3.17

🤖 Generated with Claude Code

maxconwa and others added 30 commits October 31, 2026 11:49
…y fixtures; numpy is a run-time dependency

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…dy IK path and the old controller's parity fixtures

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…ody_controller_node's policy; fix lower-body, lmg and MJPC text

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…t and the unread sim argument; fix golem_slam and bringup text

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…vo tolerances; model_server: drop the no-op approach filter; fix text

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…refusal, compose passes the H1-2 spawn and lidar-filter knobs, the Mac skips the HOMIE policy check; drop the HAMS guard and dead sim helpers

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…ontainer, G1 navigation, lower body, skill timeouts)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
Each robot's IK and held joints, frames, named postures, start-up sequence and
controller-config defaults are config/<robot>/robot.yaml, and its safety/controller
config pairs are config/<robot>/pairing.yaml. profile.known_robots() and
profile.robot_config() read them, and pairing applies one rule to every robot:
the safety config is one of the pairs, the controller config is the paired one,
and with sim given neither is on the wrong side of it. The H1-2 now refuses a
safety config outside its pairs and sim/real mixes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…_CONFIGS

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
golem_safety's safety.launch.py starts safety_node after start_delay;
golem_upper_ik's upper_ik.launch.py checks the safety/controller pairing
(upper_ik: not starting: ...) and then starts frame_task_server after
start_delay; its new robot_state.launch.py starts robot_state_publisher and
joint_state_publisher. Each resolves its arguments when included, since a
Humble TimerAction evaluates its actions when it fires, after later unscoped
includes may have reset robot, config or sim.

h12_state, h12_control and g1_core include them, upper_ik.launch.py first,
keeping their delays and node names. g1_core keeps only locomotion_refusal;
h12_control's check_pairing is gone. The bringup tests walk the includes as
ros2 launch runs them; the G1 sim tests also recognise the upper_ik refusal.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
The robocasa service interpolates it, so run.sh refused to start until it
was pinned or allowed. Pinned empty, it gives the sim's default (the robot's
own limbs dropped from the lidar), which the golden describes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
CL_Assets moves to shared_robot_stack (d7ef401): livox_link and
livox_imu_link are the sensor's raw frames, turned 180 deg about x from
the upright ones, in every H1-2 URDF and both MJCFs. h12_sensors casts its
ray fan over the MID-360's own -7..+52 deg elevation in that frame
(lidar_fan), so the sim's points and IMU come out in the raw frames with no
rotation in code.

test_mid360_frames checks every H1-2 URDF variant against the upright
frames turned about x and the IMU upside down in the pelvis;
tests/robocasa/test_h12_lidar_mount.py checks both MJCFs against the URDF
and that the fan reaches the floor and above the horizon.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…u_upright deleted

The H1-2's MID-360 hangs upside down and its points and IMU are both in the
sensor's raw frame (the models' livox_link and livox_imu_link). The real
driver's extrinsic is zero as the G1's, /livox/imu comes straight from the
driver, and the RoboCasa sim stamps both streams livox_frame, as the driver
does. The H1-2 navigation graph and the Mac bringup hang livox_frame off
livox_link with an identity static transform.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
navigation.launch.py (robot required) builds the same nodes, executables
and remaps on the H1-2 and the G1; each robot's values live in
config/<robot>/: FAST-LIO, the MID-360 driver config, Nav2, OctoMap and
pipeline.yaml (lidar link, scan band, level_odom's floor depth).
slam_toolbox.yaml is shared.

- body_to_pelvis (was g1_body_to_pelvis) runs on both robots, replacing
  the H1-2's static body -> pelvis; level_odom always takes body -> pelvis
  at each odometry stamp.
- Both robots get the static <lidar link> -> livox_frame, level_cloud and
  the scan on /cloud_base_filtered, and the behavior_server cmd_vel remap
  through the velocity smoother.
- The H1-2's FAST-LIO extrinsic (livox_link in livox_imu_link) is written
  in config/h12/fastlio.yaml; the launch no longer reads the URDF.
- Every includer passes robot; the G1 sim passes config/g1/fastlio_sim.yaml.
- golem_slam's setup.py skips dangling links a --symlink-install build
  directory keeps for deleted files.
- tests/g1: the launch-log parser strips the colour codes emulate_tty adds.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
A link with collision geometry that no TF has ever reached (the G1's Dex3
finger links with use_hands:=false) is skipped and logged once, again only
when the set changes. A link that had a TF and loses it still drops the cloud.
test_sim_nav.sh takes NAV_USE_HANDS to pass use_hands to the bringup.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
Until a grace period (default 2 s of cloud-stamp time) has passed since the
first cloud in which any link had a TF, a cloud with a link that never had one
is dropped, so start-up clouds never carry the body into the maps.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
The test scans golem_safety, golem_upper_ik and golem_slam for the profiles'
robot names in string literals, robot comparisons and file names. It found
h12_real_slam.launch.py, an H1-2 entry point in golem_slam: it moves to
golem_bringup's launch/h12/. The robot lists in the launch arguments, the
parameter docs and safety_node's missing-robot message come from the
profiles instead of naming h12 and g1.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
h12_sim_mac.launch.py includes h12_state, h12_control and, by its GOLEM_*
toggles, h12_lowerbody and golem_slam's navigation (robot h12), each with
the arguments h12_sim.launch.py passes. An unknown GOLEM_LOWERBODY is
refused. Its own nodes, the duplicate livox_frame TF, the raw-cloud scan,
the ground-truth-odom SLAM and nav2_config_mac.yaml are gone.

The slim arm64 ROS image adds PCL, pcl_ros, APR, octomap_server and
Livox-SDK2 at the x86 images' commit, and bakes livox_ros_driver2 at
/opt/livox_ws with an import check; launch_ros_mac.sh builds fast_lio
against it. rob_odom reads FAST-LIO's /Odometry. A source-pin test keeps
every image's Livox-SDK2 on one commit.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…state every physics step

The H1-2 sim's Unitree DDS command path now matches the G1 sim's. An
rt/lowcmd frame is dropped for a bad CRC, a mode_machine other than 6, a
mode_pr other than 0 or a value that is not finite on the 27 body motors,
with a warning at most once per reason per 5 s. A valid frame is latched
whole as one immutable command stamped with sim time; the DDS handler no
longer writes into shared arrays or reads MjData. Every motor is off until
the first valid frame; motor mode 1 stays the motor-on rule and no torque
clip is added.

GOLEM_CMD_TIMEOUT must be finite and positive (empty means 0.5) and is
checked in the physics step against the command's sim time, once commands
have started, with one log line on silence and one on return; the watchdog
thread is gone. rt/lowstate is published from the main loop after every
mj_step (the wall-clock publisher thread is gone) with a millisecond
sim-time tick, mode_pr 0, rpy from the IMU quaternion and a CRC. The sim
closes its channels on exit.

The helpers both robots' interfaces use (timeout knob, command latching,
rate-limited warnings, fast CRC, rpy, DDS domain) move to unitree_common.py;
the G1's behaviour is unchanged (its CRC marker is renamed _sim_fast).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…call on both leg nodes

- h12_lowerbody_rl: LegSupervisor (supervisor.py, no ROS) holds the leg
  node's phases (waiting, ramping, holding, running, faulted) and g1_homie's
  checks: upright to start, a fall, a stale or non-finite /lowstate or a
  non-unit quaternion, a false or silent heartbeat, a policy that raises or
  gives non-finite targets latch it silent (the node stays up); /cmd_vel
  times out to zero (cmd_vel_timeout), non-finite commands are zero; the
  operator confirm is accepted only while holding. New parameters
  max_state_gap (sim 0.5, real 0.05) and cmd_vel_timeout (0.5).
- Both leg nodes call /elastic_band/toggle once, when the service is ready;
  the reply is final (a failure is logged as "elastic band release failed"
  and the band stays on). The H1-2 node resets the policy's history on the
  success, not when the call is sent.
- g1_homie: in the sim the policy's start waits for the first /left_ee_pose
  (bounded by band.max_wait from the default pose), so HOMIE runs under the
  band for about one 0.1 s poll; 1.5-3 s there tilted the G1 21-27 deg at
  the release. Walk test 22/22 at 7.4 deg, nav 13/13 at 7.4 deg.
- h12_lowerbody_rl's linters run again: reference/ is excluded, the
  package's own flake8/pep257 findings are fixed, and the mypy test skips
  where ament_mypy is not installed (the ros image).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…r stalls

The leg node measures the /lowstate gap between callbacks on its one
executor, which also runs the policy tick, so the gap includes executor
stalls; lowerbody_real.yaml's max_state_gap is 0.25 s, which catches a dead
stream without latching the legs silent on a stall. CLAUDE.md/AGENTS.md §9
and the README add the pre-flight: measure the worst /lowstate callback gap
on the onboard PC in a hoisted ALMI run.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
marl_deploy_node publishes nothing, and starts no pre-pose ramp, until
golem_safety's first true /safety/heartbeat; ~/state reads
waiting_heartbeat meanwhile. The ramp starts from the pose measured on
the first tick after the beat.

test_urdf_mount_is_the_livox_pitch is xfail: the H1-2 model's
livox_imu_link is the raw MID-360 frame and lmg_deploy's IMU code is not
updated yet. The policy configs name golem_slam's fastlio.yaml.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
The arms are halted until the first true /safety/heartbeat, for good after a
false one, and while no beat has come for heartbeat_timeout (0.5 s). A halt
cancels the frame_task or named_config goal in flight on the server, so the
skill fails with the halt reason. Grippers, navigation and frontier_explore do
not depend on it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…or both robots

tests/sim/test_sim_stack.sh --h12|--g1 runs the robot's sim bringup with
navigation in RoboCasa and requires, in one run: the robot standing
unsupported after the band's release, a Nav2 goal 1.5 m ahead reached by
walking, /named_config t_pose then home succeeding while it walks, SLAM
healthy throughout (/Odometry rate, odom -> base_footprint, a growing /map,
/converted_scan gaps) and safety_node the only /lowcmd publisher. The
containers' harness moves from tests/g1/common.sh to tests/sim/common.sh,
which the G1 tests keep using; the H1-2 sim gets its own knobs and msgs_ws.

CLAUDE.md/AGENTS.md: the closed-loop checks, golem_slam as shared infra,
lmg_deploy's and Isaac's lidar frames, the verified-on line.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
controller.threshold_joint is only the success test of a held goal. The
G1's arm PD has no integral term and settles about 1e-3 rad off under
gravity, so at 0.001 rad its /named_config goals ended with success false
at the pose. Both G1 configs (sim_split, g1_split) now use 0.01 rad,
golem_upper_ik's default and the H1-2's value. The docs that called that
failure expected describe the new test; tests/sim/README.md states the
arm protocol exactly.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…ve-pose writer

tests/g1/test_sim_nav.py holds SimDriver (the drive's node, /clock, marks,
TF, spin_for, wait_future, wait_band_released, goal_in_map, close),
ContactLog and write_live_pose; its Driver and NavRecorder and
tests/sim/test_sim_stack.py's Driver and Recorder use them instead of
copies. Both tests behave as before (contacts.json also lists the contact
spans for the nav and explore tests). tests/g1/README.md names tests/sim's
dependency on these helpers.

CLAUDE.md/AGENTS.md: the closed-loop tests run the sim with ground truth
and the experiment surface off, and refuse GOLEM_SIM_ODOM turned on.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…pairing and stale-install docs

lmg_real feeds lmg_deploy /livox/imu, which it reads as upright, while the
H1-2's MID-360 publishes it raw (upside down): the launch now stops with an
OpaqueFunction before any node, and a test asserts it (lmg_sim is not
guarded). lmg_deploy's comments and README describe the IMU frame as it is.

The real H1-2 navigation graph has not run on the robot: CLAUDE.md/AGENTS.md
section 9, golem_slam's README and h12_real_robot.launch.py name the bag
replay to do first. The h12_control and h12_safety_grippers argument
descriptions state which safety/controller configs start; the stale-install
pitfall lists what an old golem_slam and golem_bringup install keeps.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
maxconwa and others added 28 commits November 9, 2026 14:14
- tests/g1/common.sh, eleven one-line forwarders, goes: the G1 scripts source
  tests/sim/common.sh and call sim_test_* as the g1 suite; test_workflows.py
  runs sim_test_setup for the G1 directly.
- test_sim_bringup.py holds the sim role's hook runner (run_recorded_sim),
  the sensors.json writer and a recorder base (StepRecorder: steps, the band's
  changes, the stop file, sampling, the atomic sim.npz) that the bringup, nav
  and stack recorders build on; test_sim_nav.spawn_grid is the occupancy grid
  both scene maps used. The stack test reads the band from SimInterface.band.
- dds_fakes.py is the fake Unitree DDS layer the three H1-2 interface and
  writer tests each wrote; spawn_scene.import_h12_mujoco the gripper-bridge
  stand-in three tests each wrote.
- The seven G1 unit files run under unittest (g1/function_tests.py collects
  their test_* functions), instead of seven copies of a hand-written runner;
  run_in_container.sh runs them as before, and the H1-2 runner can run the ones
  that open no DDS.
- test_lowstate_writer_is_faster asserts the numpy path is at least 3x faster
  (it measures ~11x) instead of only printing.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…te probe

- One IK result class: IKAttemptResult was a field-for-field copy of IKResult;
  solve_ik_reduced returns the best attempt as built.
- IKSolver.remove_frame_task folded into clear_frame_tasks, its one caller.
- RobotDynamics: get_frame_jacobian's q/imu_quat and get_gravity_compensation's
  imu_quat (no caller passed them), and its q default (every caller passes q).
- RobotModel: update_kinematics' imu_quat, the free-flyer
  get_frame_transformation (tests only; they use get_frame_transformation_body),
  and the measured-state branches of get_frame_position/rotation and
  get_frame_transformation_reduced (every caller passes q).
- The planner's hasattr/getattr guards that existed only for its test fake;
  the fake gains get_frame_rotation (identity).
- RobotClock.sim_time (tests only).
- The G1 configs' controller.imu_frame lines, which restated the profile's.
- action_msgs is a test dependency only.
- tools/publish_rate_probe.py and test_publish_rate.py measured rclpy, not this
  package; command_publisher_rate.py and frame_task_server_rate.py cover the
  rate on the real code.
- README: the Keys table points at example.controller.yaml.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…te debug.yaml

joint_state_publisher loaded and validated a whole controller config (the
robot's default_config) only to read topics.low_state. It now takes a
low_state_topic parameter, default /lowstate, so both robots' bringups (which
pass only robot) subscribe where they did. With that, nothing needs
default_config: the key leaves robot.yaml, its examples and the profile loader,
and the H1-2's debug.yaml, a byte copy of safety_full.yaml in no pair, is
deleted. frame_task_server now requires config, which every launch passes.

setup.py skips dangling links a --symlink-install build directory keeps to
deleted configs (as golem_slam's does), so existing workspaces still build.
golem_bringup's g1_core.launch.py docstring no longer says
joint_state_publisher reads g1_split.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
/frame_task and /named_config each had a direct and a planned loop: step or
hold, cancel check, feedback, convergence, sleep to the next period. One
_run_goal now runs all four, _planned_goal does the plan, execute and abort on
failure, and _move holds the slow-mode set and restore; a _GoalKind gives each
action's result type, error, feedback and threshold. The paths' differences are
kept as they were: a direct goal checks for a cancel after its step and a
planned one before; a planned frame goal ends only on its frames' errors, not
on the hold step's own convergence; a direct named goal logs "Named config
reached".

test_goal_loops.py drives every path against a scripted controller and robot
clock (cancel, timeouts, a hold blocked at a limit, plan and execution failure,
stale state, slow mode, an unknown name). Its full traces over 23,048 scripted
scenarios are identical before and after this change.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
- golem_safety_dump_config's own quoted() was a weaker copy of json.hpp's
  json_string, which the tool already links (golem_safety_core). Its output is
  byte-identical for all 11 configs. safety_node is untouched.
- README: the Keys table and the per-robot list of config values copied what
  config/example.yaml and the configs own; they point there now.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
The launch file repeated DEFAULT_KP, DEFAULT_KD and DEFAULT_MAX_SPEED as
strings; it now imports them from dex_3.dex3, and a test checks the defaults
match.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…ent the cleanup

A --symlink-install workspace built before debug.yaml was deleted keeps a
dangling share/golem_upper_ik/config/h12/debug.yaml link, which Path.glob
returns, so test_every_installed_config_is_tested failed there. It now skips
paths that do not exist, and a test checks that a dangling link is ignored.

CLAUDE.md and AGENTS.md: golem_upper_ik's setup.py, like golem_slam's, skips
dangling links, and the stale-install pitfall names the leftover debug.yaml
links and the command that deletes them. test_goal_loops.server_of's docstring
says it mirrors the attributes the goal paths read.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
- reference/ held the two original standalone deploy scripts (WALK's and
  FAME's), which nothing installs, imports or runs: they need the training
  tree's legged_gym/RMA packages. The walk and FAME math lives in
  config/h12/policies/{walk,fame}.yaml and policy/, pinned by the goldens. The
  lint exclusions and the README line for the directory go with it.
- BandRelease.in_flight was read only by test_band.py; the tests now prove the
  call in flight through result(), which takes a reply only while one is.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
- CONFIRM_BEFORE_CONTACT (a blocking input() at the pre-grasp) and
  APPROACH_IMAGE_ONLY (end at the pre-grasp and save one hand-camera frame)
  were source constants pinned to False that no parameter, launch file, test
  or doc sets. Their branches, _save_approach_image and APPROACH_IMAGE_WAIT_SEC
  go, and with them base.py's hand-camera colour cache that only the snapshot
  read (ARM_COLOR_IMAGE_TOPICS, its two subscriptions, _on_arm_image,
  latest_arm_image). The servo still reads the hand cameras through
  yolo_server's detection bundles. With both switches off the skill runs as
  before.
- grasp._quat_angle computed perception_utils.quat_geodesic's angle a second
  way; the straight-line move uses quat_geodesic (bit-identical on 200,000
  random pairs).
- _Run's docstring example is the grasp skill, not a stub.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
golem_skills.perception_utils.mat_to_quat and graspgen_server.mat_to_quat were
the same function line for line. golem_skills already depends on model_server
(it imports model_server.model_logging), so the one copy lives in a small
model_server.rotations module that both import; graspgen_server's GraspGenX
imports stay out of golem_skills. test_perception_utils.py covers it: every
branch of Shepperd's method round-trips through quat_to_rot, poses round-trip
through matrix_to_pose, and quat_geodesic is the rotation angle.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…tests

- golem_bringup.include.include replaces the eight _include copies in the
  launch files and the inline IncludeLaunchDescription(PythonLaunchDescription
  Source(os.path.join(...))) of g1_sim, g1_real_robot, h12_safety_grippers,
  h12_state and h12_real_drivers. h12_lowerbody's include is left for the
  MJPC retirement. A dump of every launch file's includes (location,
  arguments, condition), declared arguments and nodes is identical before and
  after, apart from the next item.
- h12_real_robot, h12_real_slam and lmg_real no longer pass use_sim_time to
  h12_real_drivers, which declares and reads none; the includes after it set
  it, and test_sim_time still finds every real node on the wall clock.
- main.rviz: CloudEffected (/cloud_effected, off in every FAST-LIO config) and
  LookaheadPoint/LookaheadCollisionArc (Regulated Pure Pursuit's topics;
  golem_slam runs MPPI) showed topics nothing publishes.
- Tests that proved only that removed code stays removed are gone
  (test_no_walking_node_arguments, the locomotion_node absences,
  test_one_slam_config_set); test_sim_nav_without_use_locomotion keeps its
  wiring checks as test_sim_nav_with_the_legs.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
MuJoCo MPC is retired. policy:=mjpc is now refused like any policy the
H1-2 does not list (leg_check and lowerbody.launch.py), so leg_node_condition
is start_position_verified alone. use_estimator, estimator_params and
deploy_mjpc's estimator_node go, with the MJPC configs that were all of
config/h12/ (mjpc_real.yaml, mjpc_sim.yaml, example.mjpc.yaml), their
install rule and test_example_configs.py. h12_lowerbody.launch.py includes
lowerbody.launch.py through golem_bringup.include. The bringup tests cover
policy:=mjpc's refusal in the leg block and both H1-2 bringups.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
MuJoCo MPC is retired. The ros image loses the MJPC block (clang-13/llvm/
lld/ninja, the fork's clone at MJPC_REF, grpcio/grpcio-tools/protobuf and
the mujoco==3.2.3 pin, agent_server and its libmujoco, the 2.4 GB
/opt/mjpc-build-seed and the second setuptools clamp), the Unitree C++ SDK
(/opt/unitree_install, only deploy_mjpc linked it) and python3-evdev (only
the fork's wasd_teleop.py read it). The ros image now keeps golem_base's
mujoco 3.13.0, which nothing in core_ws imports. Lock steps ros-amd64-11
and -12 go, the onnxruntime step becomes ros-amd64-11, and the common set
drops grpcio, grpcio-tools, mujoco-mpc and evdev and pins mujoco 3.13.0.

docker_build.sh no longer resolves MJPC_REF; compose drops the MJPC_REF build
arg and the three mujoco_mpc mounts; docker_common.sh no longer checks the
submodule or makes container_cache/mjpc_build; launch_ros.sh no longer
requires the Unitree SDK or hydrates and rebuilds MJPC. tests/docker loses
the two MJPC cache tests and the mujoco_mpc fixtures.

The deletions of deploy_mjpc, the mujoco_mpc submodule (with its .gitmodules
entry), rebuild_mjpc.sh and the old lock steps went in with 3fc9f69, whose
message names only golem_bringup; that commit also carried AGENTS.md's
symlink and core_ws/tf_tree.pdf's deletion (docs commit below).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…rfiles

launch_isaac.sh's HEADLESS=1 knob, which nothing sets (the wrappers pass
--headless), and LIBGL_ALWAYS_INDIRECT=0, the unset default, on the isaac
and robocasa services go. BUILD.md drops the MJPC dev loop (§7; later
sections renumbered), the package lists copied from the Dockerfiles, the
stale US apt mirror and "robosuite from master", and the
CL_isaaclab_sim/.isaac_cache migration note. SOURCES.md drops the dated
evidence-image table and the history of the old images, keeps the pin
table and policy, and names the lock steps after MJPC's removal.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
… at their sources

CLAUDE.md drops mujoco_mpc and deploy_mjpc (§2, §4: 21 packages in 18
directories, §5), the MJPC_REF note (§3), the estimator and policy:=mjpc
(§6), the 3.2.3 MuJoCo pin and the mujoco_mpc bump rule (§8, the comment
example included) and the MJPC iteration pitfall (§10); the stale-install
cleanup gains deploy_mjpc and golem_bringup's deleted config/h12/. Its
title names both files: AGENTS.md is now a symlink to it (3fc9f69), so
the two cannot drift; every reader opens it as a file.

The docs point at the sources instead of copying them: RUNNING.md's sim
topic list (CLAUDE.md §8, the surface goldens) and bringup table (§6),
g1/dds.md's test domains (the tests' READMEs) and its UDP and host-shell
sections (DDS_TUNING.md, RUNNING.md), g1/packages.md's per-package summaries
(the READMEs) and README.md's layout and documentation tables (CLAUDE.md §2,
docs/index.md). Fixed: g1/exploration.md's anchor (#validation-status),
SETUP.md's "two-machine smoke test" (there is none), DOCKER_TESTING.md's
history of an old Isaac fix, g1/index.md's port history. mkdocs.yml drops
the markdown extensions and the code-annotate feature no page uses. The
golem_lowerbody_rl README no longer offers policy:=mjpc. core_ws/tf_tree.pdf,
referenced by nothing, was deleted in 3fc9f69.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…est deps

lmg_deploy's imu_mount_quat docstring and test_imu_trim named
golem_lowerbody_rl's imu_offset_quat, which is now
golem_lowerbody_rl.imu.trim_quaternion (it returns the bias's conjugate);
its README no longer mentions an MJPC controller. No behaviour changes.
model_server has no tests, so its extras_require['test'] and
<test_depend>python3-pytest</test_depend> go.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
The 2026-11-09 cleanup does not edit the submodules (CLAUDE.md §5): audit D's
submodule findings (custom_ros_messages, magpie_msgs, magpie_control,
cl_realsense, estop, CL_isaaclab_sim) and audit A's RealtimePacer copy are
listed with their evidence, for each owner to decide.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…passes

../CLAUDE.md is outside docs_dir, so mkdocs --strict (the docs workflow)
refused the link; the other pages link repo files by their GitHub URL.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
The MJPC block's second clamp also undid the GraspGenX dependency step
(ros-amd64-08), which installs setuptools 84.2.1 and its
distutils-precedence.pth; without it the rebuilt image had setuptools 84,
which breaks colcon's ament_python builds. The clamp is now the last step,
through step 6's lock (it installs only the three pinned names), and asserts
setuptools 59.6.0 and numpy 1.x.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
sim_main.py's `import tasks` registers the Isaac Lab tasks parse_env_cfg looks
up, so S8 no longer lists it as dead and warns that pyflakes flags such
side-effect imports. The `tests/` directory it described is local and
untracked. S6 names cl_realsense/scripts/another_node.py and the launch files
the bringups use instead of h12_rs_cams.launch.py.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…OMIE imports MuJoCo

CLAUDE.md §4 names h12_headcamera.launch.py and h12_hand_cameras.launch.py,
which the bringups include, as cl_realsense's entry points instead of the
unused h12_rs_cams.launch.py; g1_head_camera.py's docstring cites the head
camera's file. §8: new pip deps in RosDockerfile go before its final clamp,
and the ros image's MuJoCo is imported by nothing the workspace runs (HOMIE's
vendored mujoco_deploy_g1.py imports it and is never run).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…os clamp last

Nothing uses OSMesa outside MJPC: in base, robocasa and ros,
`apt-cache rdepends --installed --no-suggests --no-recommends libosmesa6`
lists only Base's own libosmesa6-dev (libglfw3 merely Suggests it), and no
code or env selects it (the sims refuse MUJOCO_GL=osmesa). So golem_base
drops libosmesa6-dev; RobocasaDockerfile keeps its own libosmesa6 (a kept
package). The same rebuild deletes Base's no-op archive.ubuntu.com sed and
the rosdistro key download no apt source references.

test_source_pins keeps the setuptools/wheel/numpy clamp RosDockerfile's last
pip step; ros-amd64-06.txt, SOURCES.md and BUILD.md say the final clamp
reuses that lock and stays last, and BUILD.md that HOMIE's vendored deploy
script imports MuJoCo though nothing runs it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
Its name still pointed at the deleted g1_test_setup. The docstring now says
what the test proves: the G1 overlay on every run, the containers removed on
exit and the robot in run.json.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…l and DDS fixes

- docs/superpowers/specs/2026-11-09-repo-cleanup-leftovers.md lists every audit
  finding the cleanup kept, listed or left, by package, with where, why kept and
  the decision each owner needs. The submodule spec no longer cites gitignored
  audit files.
- README.md describes the lower body as RL locomotion; MJPC is retired.
- CLAUDE.md section 10 names container_cache/mjpc_build and
  .git/modules/mujoco_mpc as safe to delete.
- docs/g1/dds.md also points at dex_3's README for its test domain.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
It carries d7ef401, the H1-2's raw-frame MID-360, and main's manipulation
net peg-in-hole assets.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
… enabled

configure-pages failed every push to main because Pages is not enabled for the
repository. The build job now runs mkdocs build --strict on pull requests too,
and uploads and deploys the site from main only when GET /pages answers 200.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
…n flight

The recorder stamps each command when it receives it, so under load commands
published before the abort can arrive after the goal ended. The test now checks
that the held command, once it arrives, never changes, over more than 50
samples.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MezUeauFS9tKMgZKJxo1HC
@maxconwa
maxconwa merged commit ce9d6ba into main Nov 10, 2026
4 checks passed
@maxconwa
maxconwa deleted the repo_cleanup branch November 15, 2026 23:37
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