[hhhappyshow]

Summary

• TorchEnv.advance_physics(ctrl): new public override hook so task owners can drive one control interval with an already validated control (e.g. interleaving a converted low-rate ONNX low-level policy inside a longer control interval). Owners advance exactly sim_substeps via public backend APIs and may refresh the control between chunks; rewards, episode counters, and resets stay in the outer step lifecycle. The default delegates to _step_physics, so the substep feedback machinery merged earlier keeps working unchanged.
• sim2sim contract: playback-visualization-only action fields (log_goal, show_rays) are now exempt from the env.actions contract comparison. Toggling them for interactive play is not a policy-defining change. The contract and tests/training/test_sim2sim_resolver.py are updated together.
• play_interactive: interactive.sim2sim=false skips the sim2sim preflight for external checkpoints (converted from other training stacks) that have no UniLab run_config.json snapshot; a task-owned play_key_hook consumes teleop keys (arrows/Enter) before global bindings.
• visualization: adds the line debug primitive kind (plain straight segment via mjv_connector) for ray overlays; the offline renderer has no line channel, so this is UniLab-side only.
• conf/ppo/config.yaml: adds the interactive.sim2sim: true default.

Motivation: these are the UniLab-core pieces needed to host the perceptive navigation playback flow (task owners land in a downstream repository); no task code or assets are included here.

Linked Work

• Issue: none (no driving issue for this support change)
• Base branch: main

Validation

• Focused tests passed on the final local head
• make test-all — 40 pre-existing environment failures verified identical on a clean origin/main checkout (see below)

Commands actually run:

make check                                                     # Success: no issues found in 212 source files
uv run pytest tests/training/test_sim2sim_resolver.py tests/base/test_torch_env.py tests/visualization/test_interactive_playback.py -q   # 144 passed
make test-all                                                  # 40 failed, 2893 passed, 61 skipped
# control experiment on a clean detached origin/main head:
uv run pytest "tests/base/test_sim_backend_smoke.py::test_mujoco_backend_smoke_contract" -x -q   # 1 failed (same failure)

make test-all fails with 40 MuJoCo-runtime / asset-loading / CUDA-IPC errors (OSError: [Errno 9] Bad file descriptor, asset Error opening file, CUDA IPC teardown). The same representative test fails identically on a clean origin/main head, so these failures are environmental (asset hub / sandbox restrictions), not introduced by this change. Remote CI on main is the authoritative gate.

Remote CI route:

• Base main: will run on this PR's head.

Impact

• Backend impact: none (hook is backend-agnostic; no backend behavior change by default)
• Platform impact: unknown
• Training effect expected: no (play/visualization/contract-diff only; advance_physics default is behavior-preserving)

Artifacts

• W&B: none
• benchmark result: none
• video / screenshot: none
• ONNX / checkpoint: none

Checklist

• Added or updated tests where needed (test_sim2sim_resolver.py covers the visualization-field exemption)
• Updated docs if behavior or workflow changed (no public docs affected; hook docstring covers usage)
• Linked the driving issue (none exists)
• Noted any follow-up work explicitly (downstream perceptive task repository will consume the hook and play_key_hook)

Made with Cursor