Rosclaw Simforge

ROSClaw SimForge

Validate the full software loop from CLI and safety gates through simulators, MCP, evidence receipts, and asset distribution. Treat real hardware as out of scope unless the user separately authorizes it.

Safety boundary

• Use the exact rosclaw CLI from the checkout's .venv/bin; report a stale global CLI before changing it.
• Use a temporary ROSCLAW_HOME for smoke tests and Hub operations.
• Write raw trajectories, logs, receipts, and reports to an evidence directory outside the source checkout. Commit only reproducible code and tests.
• ROS graph discovery, subscription, and simulation are allowed. Never publish actuator commands directly, call a motion service, or send an action goal.
• Route any action request through rosclawd request_action / ROSClaw MCP so the daemon, sandbox, and firewall remain in the path.
• Do not stop unrelated GPU processes or containers. Inspect available memory and choose the smallest useful workload.
• A dependency failure is a failure to diagnose and install around, not a passing skip. Keep unsupported real-robot checks explicitly out of scope.

Workflow

1. Establish provenance

Read the repository AGENTS.md and the task's verification document. Record:

git rev-parse HEAD
git status --short
nvidia-smi --query-gpu=index,name,memory.used,memory.free --format=csv

Fetch or compare upstream before claiming the checkout is current. Preserve all pre-existing user changes.

2. Validate the ROSClaw core

Prepend the repository environment to PATH, use a temporary home, and run doctor, status, the deterministic demo, receipt explanation, and the universal agent/MCP probe. Then run the repository test groups requested by the project verification document.

Do not claim a simulator pass from import-only evidence. Require a physics step, a bounded task outcome, and a receipt or other machine-readable result.

3. Validate ROS 2 and Gazebo

Run scripts/verify_ros2.sh from this skill directory. It builds the Humble rosbridge/turtlesim image, deploys the stack, discovers the live graph, subscribes to pose, and proves direct velocity requests are blocked.

For Gazebo, run scripts/verify_gazebo.sh. Humble's recommended pairing is Gazebo Fortress. The script now runs the Phase 3 GuardedBase world with real diff-drive, odometry, laser, independent ROS bridges, and a deadman under launch_testing. It injects actual process signals and requires bounded stop, observation-loss fail-closed behavior, cancellation/recovery, and no old-action replay. For the full MCP → rosclawd path and rosbridge-loss recovery, run:

rosclaw chaos run gazebo-guarded-base \
  --faults agent-kill,rosbridge-loss,odom-stale,worker-crash \
  --output-dir /an/external/evidence/directory

Raw evidence must remain outside the checkout. A /clock sample alone is not enough to claim GuardedBase integration.

4. Validate Isaac Lab and four GPUs

Read references/verified-stack.md, then run scripts/verify_isaaclab.sh. The script performs bounded single-GPU and multi-GPU Cartpole training with Newton/MJWarp. Four visible devices must map to four ranks, synchronize gradients, advance physics, and exit zero.

An Isaac Sim container that advances physics but aborts during shutdown is a partial pass with a lifecycle defect, not a clean pass.

5. Validate G1 GoalForge

Use an external checkout of RoboNaldo's G1 deployment assets and keep all episode trajectories outside ROSClaw. The backend must qualify the exact 29-joint Unitree hg order, ONNX shape, motion tensors, body hash, and prior hash before physics starts.

Run the product surface from the checkout:

.venv/bin/python -m rosclaw.entrypoint simforge doctor g1-goalforge --all \
  --output /evidence/doctor/goalforge-doctor.json
.venv/bin/python -m rosclaw.entrypoint simforge validate g1-goalforge \
  --pairs 100 --output /an/external/recovery-100.json
.venv/bin/python -m rosclaw.entrypoint simforge validate g1-goalforge \
  --profile nominal-success --workers 4 \
  --output /an/external/nominal-success-30.json
.venv/bin/python -m rosclaw.entrypoint demo run g1-goalforge \
  --target-zone random --failure-to-success --live-dashboard \
  --output-dir /an/external/new/evidence/directory
.venv/bin/python -m rosclaw.entrypoint practice start \
  --task g1_penalty_kick --generation 3 \
  --output-dir /an/external/new/practice/directory
.venv/bin/python -m rosclaw.entrypoint evolution run \
  --task g1_penalty_kick --generation 10 --gpus 0,1,2,3 \
  --output-dir /an/external/new/evolution/directory
.venv/bin/python -m rosclaw.entrypoint chaos run g1-goalforge \
  --faults agent-kill,worker-crash,dds-loss,state-stale \
  --output-dir /an/external/new/chaos/directory
MUJOCO_GL=egl \
.venv/bin/python -m rosclaw.entrypoint evolution export \
  /an/external/new/evidence/directory \
  --format video \
  --output /an/external/new/video/g1-goalforge.mp4

The four-GPU screen is a prioritizer. Require a disjoint, balanced CPU MuJoCo label-agreement run before accepting its labels, preserve mismatches as counterexamples, and use CPU MuJoCo strict replay for final physical truth. Private Holdout case rows must not enter candidate generation or public output.

Build E5 proofs only from passing source reports, then independently replay their bundle hash and primitive causal/fault/replay fields:

.venv/bin/python -m rosclaw.entrypoint proof build g1-goalforge \
  --demo /evidence/demo/goalforge-demo.json \
  --recovery /evidence/recovery/recovery-100.json \
  --flywheel /evidence/practice/goalforge-flywheel.json \
  --memory /evidence/evolution/memory-ablation-100.json \
  --four-gpu /evidence/evolution \
  --agreement /evidence/evolution/cpu-gpu-label-agreement.json \
  --continual /evidence/evolution/continual-g0-g10.json \
  --chaos /evidence/chaos/goalforge-chaos.json \
  --output-dir /evidence/proofs
.venv/bin/python -m rosclaw.entrypoint proof replay /evidence/proofs \
  --modules body,provider,failure_router,sandbox,practice,memory,know,how,auto,darwin,registry,rosclawd
.venv/bin/python -m rosclaw.entrypoint promotion evaluate g1-goalforge \
  --doctor /evidence/doctor/goalforge-doctor.json \
  --recovery /evidence/recovery/recovery-100.json \
  --flywheel /evidence/practice/goalforge-flywheel.json \
  --four-gpu /evidence/evolution \
  --continual /evidence/evolution/continual-g0-g10.json \
  --chaos /evidence/chaos/goalforge-chaos.json \
  --proofs /evidence/proofs \
  --output /evidence/promotion-v4.json

The Promotion command requires a fresh simulation-only Doctor report. Treat a Doctor/Champion/Proof Body or kick-prior hash mismatch as a hard failure, and require all twelve GoalForge module proofs rather than only the four learning modules.

Unitree DDS tests must use loopback, a non-default isolated domain, canonical rt/lowcmd and rt/lowstate, simulation-only permits, and hardware_authorized=false. Never start a physical G1 transport.

GoalForge video export must consume strict-replay trajectory artifacts outside the checkout and write the MP4 and manifest outside the checkout. It is visualization-only: never use rendered pixels or subtitles as Promotion truth.

For success-rate optimization, use --profile nominal-success as the CPU MuJoCo acceptance surface. Require all 30 balanced nominal cases, at least 95% success, a gain of at least 30 percentage points over the fixed prior, 100% safe selection, 100% independent verification, and 100% strict replay. Keep the two-attempt runtime recovery budget separate from the at-most-32-candidate offline simulation search. Do not generalize that result to moving balls, randomized mass/friction/latency/noise/disturbance, or the 0.90 m target unless those cases pass a separately declared validation profile.

6. Validate MCP

First probe ROSClaw's built-in MCP server and list its tools. For the community Isaac Sim MCP adapter, isolate installation, start its extension only in a simulator, perform an MCP initialize/list-tools/call-tool round trip, and state that it is community-maintained rather than an NVIDIA-official MCP.

Never expose execute_script or scene mutation tools to real hardware.

7. Validate Hub upload/download

Run scripts/verify_hub.sh. It uses the repository fixture key and an isolated local HTTP registry to exercise validation, signature verification, login, signed publish, catalog sync/search, remote dry-run download, install, list, uninstall, and empty final state. Never reuse the fixture key for production.

8. Report the evidence ceiling

Separate PASS, PARTIAL, FAIL, and OUT OF SCOPE. Include exact commit, versions, commands, exit codes, test counts, GPU mapping, and artifact paths. The maximum claim must match the strongest verified evidence domain. Simulation evidence can promote only from baseline to SIM and never proves real-robot safety.

Resources

• scripts/verify_ros2.sh — live ROS 2/turtlesim safety loop.
• scripts/verify_gazebo.sh — Fortress diff-drive, odometry, laser, deadman, and real launch_testing process faults.
• scripts/verify_isaaclab.sh — bounded one- and multi-GPU Isaac Lab loop.
• scripts/verify_hub.sh — signed local Hub upload/download loop.
• references/verified-stack.md — tested versions, expected evidence, and known compatibility limits.