메뉴
Benchmark and analyze Moleworks IsaacLab RL checkpoints. Use when playing a policy locally, benchmarking excavation checkpoints, ranking multiple checkpoints fairly, plotting TensorBoard progression, or comparing synced run artifacts before recommending a checkpoint.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Current Moleworks Terra trenching runbook for full autonomous Beam6-style trench execution in simulation or on the robot. Use when investigating or running the two-stage flange/bottom trench flow, generate_trench_sequence_plans.py, beam6_sequence_stage.launch.py, BASE_CONTROL target registration, mesh_to_excavation_grid_map.py, workspace planner trench-axis metadata, Terra behavior-tree activation, Newton or Isaac/Terra simulation bringup, robot bringup, and 400 mm tool handoff.
Read recent Google Chat context, draft or send a reply in the correct DM or space, download collaborator attachments such as timesheets or PDFs, and handle simple meeting coordination by creating or updating a Google Calendar invite and posting the Meet link back in Chat. Use when Lorenzo asks to read a collaborator's recent messages, understand chat context before replying, send a Google Chat reply through the Chat API, pull a PDF or timesheet out of Chat, or create a meeting from a chat exchange.
Replay split DIG bags in the `moleworks_ros` container with bag TF, live self-filter, live elevation mapping, live excavation mapping, and Foxglove. Use when reviewing DIG episodes from `sensors/`, `state/`, `commands/`, `lidar/`, and optional `elevation_map/` bags.
Finalize RSL student grading and offboarding. Use when Lorenzo asks to find a student's grading sheet, extract or submit a grade, update the RSL student-project tracker like the onboarding workflow, request eDoz grade entry from admin staff, mark offboarding fields such as completed/report/grading/source/access-revoked only with evidence, or send a short Google Chat status reply after the handoff.
Validate the Newton + ROS Nav2 driving stack in a clean tmux session after bringup. Use when the user wants a repeatable navigation check in Newton sim, including health checks for the bridge/model/drive path and the lateral-shift golden test.
Start or restart the Moleworks ROS2 stack using the Newton simulator in the default moleworks_ros runtime shell, assuming the current shell is already inside the target container unless the user says otherwise. Use when you need a clean tmux layout for Newton bridge, robot/TF/RViz, perception (elevation + excavation mapping), optional Foxglove bridge, an isolated bridge-only validation stack on a specific ROS domain, or Terra failure capture and resume from saved checkpoints in Newton simulation, all with use_sim_time:=true.
| name | rl-isaaclab-benchmark |
| description | Benchmark and analyze Moleworks IsaacLab RL checkpoints. Use when playing a policy locally, benchmarking excavation checkpoints, ranking multiple checkpoints fairly, plotting TensorBoard progression, or comparing synced run artifacts before recommending a checkpoint. |
Use this skill for checkpoint evaluation, playback, and run-to-run comparison in moleworks_ext.
Read only the files needed for the current task:
docs/AI_RESEARCHER_WORKFLOW.mddocs/EXPERIMENTS_ONGOING.mddocs/EXPERIMENTS_RUN.mdscripts/rsl_rl/play.pyscripts/mole_environments/excavation3D/benchmark_excavation.pyscripts/utils/compare_run_configs.pyscripts/plot_tb_scalars.pyscripts/rsl_rl/play.pyscripts/mole_environments/excavation3D/benchmark_excavation.pyscripts/utils/compare_run_configs.pyscripts/plot_tb_scalars.py/workspace/isaaclab/isaaclab.sh -p scripts/rsl_rl/play.py \
--task <TASK> \
--checkpoint logs/rsl_rl/<exp>/<run_dir>/model_<N>.pt \
--num_envs 1
Use --num_envs 1 for visual debugging unless the user explicitly wants a larger headless check.
Excavation3D / w-cabin template:
/workspace/isaaclab/isaaclab.sh -p scripts/mole_environments/excavation3D/benchmark_excavation.py \
--task Moleworks-Isaac-m445-digging-3D-w-cabin \
--checkpoint logs/rsl_rl/<exp>/<run_dir>/model_<N>.pt \
--num_envs 2048 \
--benchmark_steps 300
For fair cross-run comparison:
num_envs fixedbenchmark_steps fixedseed fixedWhen sweeping checkpoints, use one fixed benchmark contract:
RUN_DIR=logs/rsl_rl/<exp>/<run_dir>
for m in 500 1000 1500 2000; do
/workspace/isaaclab/isaaclab.sh -p scripts/mole_environments/excavation3D/benchmark_excavation.py \
--task Moleworks-Isaac-m445-digging-3D-w-cabin \
--checkpoint "${RUN_DIR}/model_${m}.pt" \
--num_envs 1024 \
--benchmark_steps 400 \
--seed 0
done
Rank from the generated benchmark reports, not from "latest checkpoint wins".
After pulling or benchmarking a run, generate progression plots:
python3 scripts/plot_tb_scalars.py \
--run-dir logs/rsl_rl/<exp>/<run_dir> \
--plot-core \
--out-dir outputs/analysis/training_curves
Use exact tags or regex when the default core plot is not enough:
python3 scripts/plot_tb_scalars.py --run-dir logs/rsl_rl/<exp>/<run_dir> --list-tags
python3 scripts/plot_tb_scalars.py --run-dir logs/rsl_rl/<exp>/<run_dir> --regex '^Episode_Termination/'
compare_run_configs.py before manual config diffs.moleworks-subagent-orchestrator.newton-ros-parity and ros2-debugging and make sure the ROS side is cleaned up before the next IsaacLab launch.