rl-newton-benchmark

name: rl-newton-benchmark description: "Benchmark and analyze Moleworks Newton RL checkpoints. Use when benchmarking analytic cabin or shared-turn checkpoints, collecting or replaying terrain banks, verifying a saved bank, comparing fresh vs carved terrain, or summarizing benchmark JSON outputs."

Newton Benchmark Workflow

Use this skill for checkpoint evaluation, terrain-bank workflows, and benchmark interpretation.

Source Of Truth

Read only what matches the task:

• docs/PARTIALLY_DUG_TERRAIN_WORKFLOW.md
• docs/SHARED_TURN_DEFAULT_TRAINING.md
• docs/benchmarks/README.md
• docs/experiments/README.md
• scripts/mole_environments/fee_excavation/README.md
• scripts/mole_environments/fee_excavation/benchmark/README.md
• scripts/mole_environments/fee_excavation/benchmark/benchmark_fee_excavation.py
• scripts/benchmark/benchmark_excavation_w_cabin_analytic.py
• scripts/benchmark/verify_success_terrain_bank.py
• scripts/benchmark/analyze_benchmark.py

Hard Rules

• Benchmark before promoting a checkpoint or recipe.
• Run benchmark executions and noisy benchmark-log parsing in a worker agent by default. The parent thread should keep only the command contract, report paths, and compact metrics.
• For shared-turn analytic tasks:
  • use the analytic env
  • keep empirical normalization disabled
  • use the torch soil backend only
• Compare fresh-RBF performance before reasoning about carved-terrain replay.
• Do not trust a saved terrain bank until same-episode verification passes.

Main Entry Points

• Generic excavation benchmark: scripts/benchmark/benchmark_excavation.py
• Shared-turn / analytic cabin benchmark: scripts/benchmark/benchmark_excavation_w_cabin_analytic.py
• FEE excavation benchmark: scripts/mole_environments/fee_excavation/benchmark/benchmark_fee_excavation.py
• FEE category leaderboard: scripts/mole_environments/fee_excavation/benchmark/benchmark_fee_leaderboard.py
• Terrain-bank verifier: scripts/benchmark/verify_success_terrain_bank.py
• Result summarizer: scripts/benchmark/analyze_benchmark.py
• Visual sanity checks: scripts/debug/capture_terra_snapshot.py, scripts/debug/capture_shared_turn_snapshot.py

Delegated Benchmark Execution

Use a worker subagent for benchmark runs, benchmark sweeps, large JSON/report comparisons, or any task likely to print long Newton/Warp/IsaacLab logs. This is Lorenzo's standing preference for this skill unless the active turn explicitly asks to keep execution in the parent thread or delegation tools are unavailable.

Parent responsibilities:

• choose the benchmark contract: checkpoint(s), seed, curriculum level, terrain source, env count, step budget, output directory, and any reset-cache namespace
• pass only the exact commands and expected summary fields to the worker
• keep the main context free of stdout/stderr; do not paste full benchmark logs into the parent conversation
• after the worker returns, read only the final report JSON/Markdown if a specific number needs verification

Worker instructions should include:

You are the Newton benchmark worker. Run the provided benchmark commands sequentially on the local GPU; do not start concurrent benchmark processes. Keep raw stdout/stderr out of your final answer unless a command fails. Write outputs under the requested output dirs. Return only: command status, report paths, success rate, desired_full/close/partial counts, major negative terminations, key tracking precision, torque/load metrics when present, and any failure traceback summary.

If a benchmark fails, the worker should return the shortest useful failure diagnosis plus the command that failed. The parent can then decide whether to inspect logs or patch code.

FEE Benchmark Bundle

Use this as the default contract for fee_excavation checkpoint evaluation, especially for hard-soil questions.

• Do not treat live training ratios as a substitute for benchmark results. If the user asks about hard-soil robustness, tracking precision, or load behavior, read or run the pinned benchmark.
• The default readout must include:
  • overall success rate plus per-type counts or rates for desired_full, desired_close, desired_partial, and negative terminations
  • target-depth precision from the benchmark report or JSON: in-soil target-depth dwell, closest approach or best-from-above, overshoot, and penetration
  • hard-soil preset performance on the current suite: default_hard_soil_mix, fixed_soil_random_hard_cap, fixed_soil_hard_interp_25, fixed_soil_hard_interp_50, fixed_soil_hard_interp_75, and asphalt_like
  • load usage under hard soil: in_soil_tau_sat_step_mean, per-joint torque saturation, torque-stage clipping (preclip vs applied over-limit fractions), and resisting force or power
• For hard-soil failure diagnosis, use the episode JSON to cut negatives by phase, not just by reason:
  • in_soil_step_count == 0
  • in_soil_step_count in [1, 3]
  • in_soil_step_count in [4, 10]
  • in_soil_step_count >= 11
  • cross-check those bins against episode_ever_close_gate, episode_ever_high_gate, episode_ever_curl_gate, and episode_best_fill_ratio
• If most negatives happen before the fill thresholds used by finish shaping and before any high/curl/close gate, treat the problem as entry/contact handling, not as finish discovery.
• If most negatives happen after the episode has already hit high/curl/close, treat the problem as lift-out / finish / spill control instead.
• Negative attribution is ordered in the task runtime. bucket_velocity is currently attributed before bucket_angle_of_attack, so benchmark JSON alone does not tell you how often those two raw masks co-fired on the same step.
• If bucket_velocity dominates and the user asks whether the policy is simply commanding too much speed, do one more check before proposing changes: compare the last few qd_desired arm targets against measured arm_joint_vel near failure. If commanded fractions are still modest but measured arm velocity spikes or reverses sign after contact, classify that as entry/contact-control instability under load, not simple policy over-commanding.
• Use that phase split before recommending interventions:
  • mostly early bins suggest entry control, near-soil speed/AoA shaping, or hardness inference
  • mostly mid bins suggest in-soil progress / finish discovery
  • mostly late bins suggest curl / lift / pullup gating or reward issues
• When comparing checkpoints, keep the benchmark contract pinned: same checkpoint number, seed, curriculum level, env count, step budget, and hard-soil preset list.
• When summarizing a hard-soil result, report both finish behavior and tracking behavior. A checkpoint that reaches desired_close with shallow penetration is not equivalent to one that reliably reaches desired_full.
• On fixed hard-soil presets, desired_full is often impossible and is not the main score. Use it as a secondary signal only; the main score is whether the policy tracks the target depth, manages load, avoids unstable negatives, and stays productive under resistance.

FEE Category Leaderboard

For FEE profile-v2 policy comparisons, prefer the checked-in wrapper instead of manually replaying ad hoc /tmp/fee_geo14_latest_bench commands:

export WANDB_MODE=disabled
uv run python scripts/mole_environments/fee_excavation/benchmark/benchmark_fee_leaderboard.py \
  --source-results-root /tmp/fee_geo14_latest_bench \
  --output-root /tmp/fee_category_leaderboard_bench \
  --num-envs 128 \
  --benchmark-steps 180 \
  --seed 0 \
  --device cuda:0 \
  --curriculum-level 100
unset WANDB_MODE

Delegate this runner to a benchmark worker when possible. It runs child benchmarks serially, writes noisy logs to per-run files, reuses compatible reset caches across FEE namespaces, and emits fee_category_leaderboard.html. The fast default contract is 128 envs, 180 benchmark steps, and a shared reset cache of 10000 generated / 8000 required samples. Interpret the report by terrain category and tracking quality, not only aggregate success: RBF/profile, entry/exit/continuing, partial-state, 5 cm dwell, shallow fraction, overshoot fraction, best in-soil error, and torque over-limit are all part of the promotion decision.

The default FEE leaderboard suite should stay small and in-distribution for the current training geometry: default, hard_random_scrape_05cm, hard_scrape_50cm, precision_profile_10cm, precision_profile_entry45deg_10cm, and precision_profile_exit45deg_10cm. The flat hard-soil scrape core case is 50 cm below the current soil and uses fixed_soil_random_hard_cap, which is in the random-mix soil support; hard_scrape_05cm remains an optional shallow stress probe. Benchmark-only soils such as interp-25/50/75/asphalt are OOD stress probes, not default promotion cases. Broader 5cm depth/profile or 30/45/60deg sweeps are optional follow-ups, with 60deg treated as an out-of-distribution stress case, not the default promotion view.

Use the canonical aggregate latest page as the browser default:

outputs/fee_benchmarks/fee_category_leaderboard_latest.html

Refresh it without running benchmarks:

uv run python scripts/mole_environments/fee_excavation/benchmark/benchmark_fee_leaderboard.py \
  --global-latest-only

This page is the global view across saved outputs/fee_benchmarks/**/benchmark_results_*.json files and should include the current table tabs, sortable columns, policy-click filtering, and video links discovered under archived videos* folders. Per-batch leaderboards remain useful for provenance, but do not present them as the default "latest" view. Normal leaderboard runs refresh the global latest page unless --skip-global-latest is passed.

When running benchmarks or videos from a worktree, treat the worktree outputs/fee_benchmarks/<batch> directory as staging only unless --output-root already points inside the canonical archive tree. The global builder scans the canonical repo root, not arbitrary worktree output roots. Before telling the user to open the full leaderboard:

CANONICAL_BENCH_ROOT=/home/lorenzo/moleworks/moleworks_newton/outputs/fee_benchmarks
rsync -a outputs/fee_benchmarks/<BENCH_ROOT>/ "$CANONICAL_BENCH_ROOT/<BENCH_ROOT>/"
uv run python scripts/mole_environments/fee_excavation/benchmark/benchmark_fee_leaderboard.py \
  --global-latest-only
rg -n "<POLICY_LABEL>|<RUN_ALIAS_OR_UNIQUE_TOKEN>" \
  "$CANONICAL_BENCH_ROOT/global_latest/fee_category_leaderboard.html" | head

Only call the global leaderboard ready after the rg check finds the new policy rows in global_latest/fee_category_leaderboard.html. If videos were recorded, verify that the video links in the global HTML point under $CANONICAL_BENCH_ROOT/<BENCH_ROOT>/videos*, not the worktree staging path. Give the browser URL for the canonical global page by default; give the per-batch HTML only when the user explicitly asks for that batch.

For qualitative FEE videos, use record_fee_leaderboard_videos.py. Default behavior should record all raw Newton MP4s first and then batch-transcode browser-safe _h264.mp4 / _vp9.webm variants in parallel across CPU workers. Do not reintroduce per-rollout inline transcoding unless explicitly requested; it leaves the GPU idle while FFmpeg runs. Useful flags:

uv run python scripts/mole_environments/fee_excavation/benchmark/record_fee_leaderboard_videos.py \
  --benchmark-output-root outputs/fee_benchmarks/<BENCH_ROOT> \
  --policy-label <POLICY_LABEL> \
  --episodes 1 \
  --cases default,hard_random_scrape_05cm,hard_scrape_50cm,precision_profile_10cm,precision_profile_entry45deg_10cm,precision_profile_exit45deg_10cm \
  --viewer gl \
  --soil-wireframe-mode full \
  --run

Use --skip-browser-video-variants for raw-only recording, --transcode-workers <n> to cap CPU parallelism, --ffmpeg-threads-per-worker <n> to control FFmpeg threading, and --inline-browser-video-variants only for the old slower inline transcode behavior. After videos finish, archive the whole batch into the canonical benchmark tree, rebuild the global latest leaderboard, and verify the policy/video links there before reporting the full leaderboard path.

For GL recording on Lorenzo's workstation, prefer the real X display. The Codex shell may have DISPLAY unset even though a display exists; check /tmp/.X11-unix/X*, validate with DISPLAY=:<n> xdpyinfo, and run the recorder/play command as env DISPLAY=:<n> ... --viewer gl --headless. Use xvfb-run only when there is no usable real display, and do not switch to USD unless the user explicitly asks.

For torque/load readouts, always report both torque stages when present:

• in_soil_tau_preclip_*: policy/servo torque before applied compensation
• in_soil_tau_sat_* and in_soil_tau_applied_over_limit_*: actual applied torque including compensation

If the user says "do not go above 1", assume they mean applied torque until clarified. A policy can have near-zero preclip over-limit while still violating applied torque because compensation is added after clipping.

Shared-Turn Terrain-Bank Contract

For partially dug shared-turn work:

• use torch, not warp
• collect only terminal desired_full and desired_close successes
• require a nonzero applied excavation footprint before saving a bank entry
• require --enable-soil-height-updates during collection
• verify the bank against the terminal same-episode terrain before trusting replay

Verification must require:

• saved_bank_to_exported_max_abs_change == 0
• exported_to_terminal_selected_surface_max_abs_change == 0
• nonzero initial_to_terminal_*_max_abs_change

Evaluation Order

Use this order for carved-terrain work:

1. Fresh-RBF benchmark on the checkpoint.
2. Collect a success-only carved bank if needed.
3. Verify one saved bank entry with verify_success_terrain_bank.py.
4. Replay-benchmark with --terrain-bank-source terrain_bank.
5. Only if replay is materially worse, consider mixed finetuning.
6. If mixed finetuning still loses on the carved benchmark, probe pure terrain_bank finetuning.

Command Templates

Shared-turn replay benchmark:

export WANDB_MODE=disabled
uv run python scripts/benchmark/benchmark_excavation_w_cabin_analytic.py \
  --task m445_excavation_shared_turn_w_cabin \
  --checkpoint <CHECKPOINT> \
  --num-envs 1024 \
  --benchmark-steps 500 \
  --device cuda:0 \
  --soil-backend torch \
  --curriculum-level 100 \
  --disable-empirical-normalization \
  --terrain-bank-source terrain_bank \
  --terrain-bank-path <BANK_PT> \
  --output-dir outputs/benchmark_terrain_bank/<RUN_NAME>
unset WANDB_MODE

Same-episode verifier:

export WANDB_MODE=disabled
uv run python scripts/benchmark/verify_success_terrain_bank.py \
  --task m445_excavation_shared_turn_w_cabin \
  --checkpoint <CHECKPOINT> \
  --device cuda:0 \
  --soil-backend torch \
  --num-envs 16 \
  --max-policy-steps 180 \
  --output-dir outputs/terrain_bank_verification/<RUN_NAME>
unset WANDB_MODE

Context Hygiene

• For existing small result files, the parent may inspect JSON directly.
• For new benchmarks, repeated comparisons, or noisy log parsing, delegate to the benchmark worker and keep only compact result tables in the parent.
• Prefer exact benchmark commands with pinned contracts over ad-hoc reruns.
• If delegation tools are unavailable, run the minimum benchmark locally, keep output bounded, and summarize from the generated report files rather than from streamed logs.
• If the task spans many result files, W&B runs, cluster syncs, or ledger updates, also use moleworks-subagent-orchestrator.