| # Using the Two Snapshots |
|
|
| Code repository: https://github.com/NEBULIS-Lab/MimicX |
|
|
| All preparation commands below are CPU-only. Use Python 3.10+; install the |
| MimicX code and its core dependencies before gate replay. Actual rollout commands |
| require the pinned simulation/tracking environment and allocated compute. |
|
|
| ## Verify Downloaded Files |
|
|
| Run from the parent of both downloaded snapshot directories: |
|
|
| ```bash |
| python MimicX-Policies/prepare.py verify --root MimicX-Policies |
| python MimicX-Assets/prepare.py verify --root MimicX-Assets |
| ``` |
|
|
| ## Recreate the Exact Core Inputs |
|
|
| ```bash |
| python MimicX-Policies/prepare.py core-inputs --policies MimicX-Policies --assets MimicX-Assets --output paper-core-inputs |
| python /path/to/MimicX/scripts/reproduction/configure_core.py --assets paper-core-inputs --spec paper-core-inputs/manifest.json --backend /path/to/MimicX/third_party/unitree_rl_mjlab --runtime-python /path/to/tracking/bin/python --output paper-config |
| ``` |
|
|
| Supply the emitted `manifest.json` explicitly: historical parameter dumps are |
| sanitized and therefore have updated hashes. Policy and reference hashes remain |
| unchanged. This command recreates all 22 required core assets across the two |
| snapshots without duplicate weight storage. Continue with the matrix/HLoop |
| instructions in the code repository's reproduction guide. |
|
|
| ## Evaluate a Released Final Policy Without Retraining |
|
|
| ```bash |
| python MimicX-Policies/prepare.py prepare-eval --policies MimicX-Policies --assets MimicX-Assets --code /path/to/MimicX --backend /path/to/MimicX/third_party/unitree_rl_mjlab --runtime-python /path/to/tracking/bin/python --trial core__tennis__m3_full_mimicx__seed101 --output tennis-evaluation |
| ``` |
|
|
| Add `--video` to prepare 1080p video and screenshot output alongside telemetry. |
| Inspect `tennis-evaluation/evaluate.sh`, then execute that script inside your |
| compute allocation. Preparation does not execute it. All trial IDs are listed |
| in `MimicX-Policies/policy_index.json`; method names are given alongside internal |
| join IDs. Use `--external-motion` for breadth references that are not distributed. |
| The supplied file must match the registered reference hash. |
|
|
| ## Replay the Core Execution Gate on Recorded Telemetry |
|
|
| ```bash |
| python MimicX-Assets/prepare.py verify-gates --assets MimicX-Assets --code /path/to/MimicX |
| ``` |
|
|
| This temporarily decompresses the recorded per-step CSVs, recomputes rollout |
| aggregates with the released metric loader and replays the actual gate. It does |
| not simulate a robot, claim a new performance measurement or perform PPO updates. |
| Expected: all 12 historical decisions match, with six acceptances and six |
| incumbent protections. For HLoop, use the core configuration's generated fixed |
| workload and compare it with archived timing/selection records under `evidence/hloop`. |
|
|
| ## Artifact URIs |
|
|
| - `policies://path`: resolve relative to MimicX-Policies. |
| - `assets://path`: resolve relative to MimicX-Assets. |
| - `external-reference://sha256`: acquire the identified licensed motion separately. |
| - `unbundled://id/name`: historical metadata pointer to a file intentionally omitted. |
|
|
| Do not execute archived command strings or Python-tagged configuration dumps. |
| Use the portable code-repository launchers and this preparation helper. |
|
|
