Cómo usar
- Copia el comando y revisa el paquete antes de instalarlo en tu entorno de agente.
- Consulta las herramientas, credenciales y entornos necesarios en el documento original.
- Describe las entradas y el resultado deseado; contrasta la salida con las fuentes.
npx skills add https://github.com/google-deepmind/robotics-antigravity-plugin --skill sim2real-gauntletAntes de empezar
- La instalación no incluye suscripción al modelo, cuota API ni permisos sobre tus archivos.
- Las estrellas e instalaciones reflejan actividad, no una prueba independiente de calidad o seguridad.
Contenido del repositorio de origen con títulos, ejemplos, código, tablas, enlaces e imágenes preservados.
Sim-to-Real Gauntlet (robosim_gauntlet)
Library and CLI in src/robosim_gauntlet (put src/ on sys.path; cli.ensure_gl_backend() preserves explicit MUJOCO_GL or selects packaged egl/osmesa). Reference consumers: playground_tasks.py and color_blob_ik_controller.py.
Workflow
- Task (
task.Taskprotocol):build_device_layer(mode="sim"|"real", extra_devices)returns(numpy DeviceLayer, reset, close)(Impl.CPPin sim; real homes onreset()and parks onclose()). Alsodeployment_keys(),success(poses, poses0), andsafety(SafetyLimits: 3D contact wrenches + indirect crush, 35 N grasp cap, bimanualarm_prefixes/gripper_tokens/command_to_joint, anddrop_z_tol_m=None/require_release=False/goal_setterfor insertions, hinged fixtures, or tool-in-hand tasks). - Controller: either an out-of-process command driving
proxy.RemoteDeviceLayer(os.environ["GAUNTLET_URL"])(prompt inGAUNTLET_PROMPT, printing{"status": "done"|"abort", "fault": ...}last) or an in-process callablefn(dl) -> dictwrapped inwrappers.AllowlistWrapper(include_latency=Falsedefault). Both see onlydeployment_keys()+<cam>_timestamp(upstream timestamps forwarded/delayed; camera reads tracked vialast_cam_read_tick). - Preflight:
preflight.run(task)audits scene invariants + negative/positive goal controls (safety.goal_setterfor articulated fixtures). - Sweep / CLI:
episode.run_suite([task], cmd_or_fn, episodes=..., mcap_dir=...)orpython -m robosim_gauntlet.cli --task-py=path.py[:fn] --robot=yam --controller="..." --output-json=v1.json --baseline-json=v0.json --mode=sim|real --mcap-log-dir=...(setGAUNTLET_CRN_SECRETfor HMAC CRN). - Debug:
episode.fork(result, task, tick, cmd_or_fn)replays logged commands totickdeterministically, then hands off;McapWrapperwrites.mcaplogs viamujoco_playground.real.create_loggeridentically insimandreal.
Tiers
| Tier | Perturbation |
|---|---|
| T0 | nominal |
| T1 / T2 | moderate / adversarial pose, table, mass, friction, gains, noise, hue |
| T3 | slip, mid-approach push, or post-place knockout (idx % 3) |
| T4 | self-model mismatch + 50% terminal fault: camera_freeze (frozen or restamped) or dead_gripper |
T4 faults must end in {"status": "abort", "fault": <fault>}; aborting without a fault is spurious_abort. Verdicts: BETTER / WORSE / INDISTINGUISHABLE (one-sided exact McNemar, alpha 0.025, >= 6 discordant pairs) with safety veto.