feat(humanoid-walk): holosoma-parity g1-walk-flat recipe with reset-time dynamics randomization

[wlgys8]

Summary

Align the g1-walk-flat fastsac recipe with holosoma's g1_29dof_fast_sac configuration. Both nightly tracking gates pass under the full randomization stack: lin-vel exp 0.522 (gate >= 0.475), ang-vel exp 0.660 (gate >= 0.533) at 30k iterations.

Recipe changes

• sim dt 0.01 (100 Hz physics; retained for the validated recipe and performance budget)
• position action scale 0.5 → 0.25 — larger scales caused per-step joint-target churn (~0.7 rad/step at 0.8) that suppressed yaw tracking (exp 0.30 → 0.73 across the sweep)
• alpha_init 0.001 / target_entropy_ratio 0.0, 30k iteration budget

Domain randomization (holosoma parity, push excluded)

• New WalkRandomizationCfg with independently disable-able items, sampled in-kernel at reset through the sim write program:
  • kp/damping ×U(0.9, 1.1), ground friction ×U(0.5, 1.25)
  • non-base link mass ×U(0.9, 1.2), base mass +U(-1, 3) kg, base com ±0.05 m
• New BodyMassesQuery (core contract + MotrixSim backend) returning a body's nominal link masses in body order
• Dedicated randomized reset kernel: numba Map getitem is statically compiled, so enabled configs declare a stable write-key schema; degenerate ranges perform no randomized writes
• WbtJointPositionActionCfg.action_delay_steps: per-lane control-step delay queue (resampled at reset); observations and the action-rate penalty keep tracking the raw undelayed actions
• WalkCommandCfg.gait_period_randomization_width: per-episode gait-period jitter (±0.2 s) via a per-lane phase_dt buffer
• BadDofVelocityTerminationCfg: reusable core termination shared by WBT and other manager tasks; resets lanes whose joint speeds exceed any healthy gait after harsh impacts

Results (30k iters, 25–30k window)

metric	no DR	+ full DR	holosoma gate
lin-vel exp	0.599	0.522	≥ 0.475 ✅
ang-vel exp	0.725	0.660	≥ 0.533 ✅
mean ep_len	1000	991	—

Training is stable under the joint perturbation stack (near-full episodes, zero collapses).

Test plan

• new test_walk_randomization.py (cfg validation, delay behavior, gait band, query wiring)
• existing test_humanoid_walk.py / test_wbt_numba.py suites pass unchanged
• 30k end-to-end training run with gates verified above

[Copilot]

Copilot review overview

🔵 Needs a closer look

It spans a core sim-API rename, numba-kernel compilation semantics (a latent build failure for partially-disabled randomization configs), and an RL recipe change whose committed dt contradicts the PR description, warranting human review.

Review effort: Balanced
Findings: 2 · 2

Open (4)

• g1-walk-flat still overrides sim dt to 0.01 · New
• Randomized reset kernel references missing sim_writes keys · New
• New dof-velocity termination is unused and duplicated · New
• Test pins recipe tuning values too tightly · New

What changed in this PR

This PR aligns the g1-walk-flat FastSAC recipe with holosoma's g1_29dof_fast_sac configuration by adding reset-time dynamics randomization. It introduces a new WalkRandomizationCfg (kp/damping/friction/mass/base-com randomized in-kernel at reset via the sim write program), per-episode gait-period jitter, a per-lane action-delay queue, and a BadDofVelocityTerminationCfg. Supporting this, it renames the single-link sim queries/writes to a Link* namespace and adds a tree-scoped BodyMassesQuery to the backend-neutral sim contract.

Changes:

• Rename BodyMass*/BodyCom*/BodyCenterOfMass* queries & writes to Link*, add BodyMassesQuery (core contract + MotrixSim backend) returning a body tree's nominal link masses in tree order.
• Add reset-time domain randomization (WalkRandomizationCfg + dedicated randomized reset kernel), gait-period jitter (WalkCommandCfg.gait_period_randomization_width), action-delay config, and BadDofVelocityTerminationCfg.
• Retune the g1-walk-flat recipe/config (action scale 0.25, alpha/entropy, 30k iterations) and add test_walk_randomization.py.

File	Description
motrix_env_core/​src/​.../​sim/​model.py	Rename to LinkMassQuery/LinkCenterOfMassQuery, add BodyMassesQuery + abstract compile_body_masses.
motrix_env_core/​src/​.../​sim/​write.py	Rename BodyMassWrite/BodyComWrite → LinkMassWrite/LinkComWrite; document Link/Body naming.
motrix_env_core/​src/​.../​sim/​__init__.py	Update exports for renamed/new query types.
motrix_env_core/​src/​.../​mdp/​terminations.py	Add BadDofVelocityTerminationCfg + kernel (not wired into any recipe; duplicates WBT name).
motrix_env_motrixsim/​src/​.../​runtime.py	Rename compile hooks; implement compile_body_masses (tree-order link masses).
motrix_env_motrixsim/​src/​.../​write_compiler.py	Rename compile_link_mass/compile_link_com.
motrix_envs/​src/​.../​humanoid/​walk_manager_mdp/​randomization.py	New WalkRandomizationCfg with range validation.
motrix_envs/​src/​.../​humanoid/​walk_manager_mdp/​reset.py	New randomized reset kernel + param assembly; conditional write declaration.
motrix_envs/​src/​.../​humanoid/​walk_manager_mdp/​command.py	Per-lane phase_step buffer with gait-period jitter.
motrix_envs/​src/​.../​humanoid/​cfg.py	Wire randomization model queries when enabled.
motrix_envs/​src/​.../​humanoid/​g1.py	Enable full randomization stack on g1-walk-flat (sim still dt=0.01).
motrix_envs/​src/​.../​quadruped/​walk_np.py	Adopt renamed Link* queries/writes.
configs/​task/​g1-walk-flat/​motrix.fastsac.yaml	Alpha/entropy parity + 30k iterations.
motrix_env_core/​tests/​*	Update dispatch tests for renamed hooks.
motrix_envs/​tests/​test_walk_randomization.py	New behavioral tests (some assert exact tuning values).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.