google-deepmind/robotics-antigravity-plugin

sim2real-iteration

Guides disciplined MuJoCo robotics sim-to-real controller iteration using robosimgauntlet verification, preflight scene/goal audits, deterministic episode forking, paired Common Random Numbers, and adversarial robotics critic reviews.

ソースを見る

使い方

  1. インストールコマンドをコピーし、自分の環境へ導入する前にソースを確認します。
  2. 必要なツール、認証情報、対応環境を原文で確認します。
  3. 入力と成果物を指定し、出力を資料と照合します。
npx skills add https://github.com/google-deepmind/robotics-antigravity-plugin --skill sim2real-iteration

開始前の確認

  • 導入にはモデル契約、API枠、ファイルへのアクセス許可は含まれません。
  • スター数や導入数は活動の指標であり、独立した品質・安全性試験ではありません。
リポジトリの原文

見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。

Sim-to-Real Controller Iteration

Use alongside sim2real-gauntlet (src/robosim_gauntlet on sys.path) when tuning or debugging a manipulation controller for physical deployment. Never claim an improvement from single-seed runs, T0-only checks, or unpaired rollouts. Reference consumers: ../sim2real-gauntlet/examples/playground_tasks.py and ../sim2real-gauntlet/examples/color_blob_ik_controller.py.

Loop

  1. Preflight gate: Audit compiled MjModel physics and positive/negative goal controls (safety.goal_setter for articulated fixtures) before iterating:
bash
   python -m robosim_gauntlet.cli --task-py=../sim2real-gauntlet/examples/playground_tasks.py --robot=yam --tasks=cube_in_bowl --preflight-only

Fix every reported issue (manipuland welds, non-robot actuators, manipuland margin/gap, contact overrides, or vacuous/unsatisfiable task.success).

  1. Unprivileged controller contract: Drive proxy.RemoteDeviceLayer(os.environ["GAUNTLET_URL"]) out of process (or an in-process fn(dl) behind wrappers.AllowlistWrapper) using only task.deployment_keys() and <cam>_timestamp. On T4 terminal faults (camera_freeze or dead_gripper), observe post-onset state and exit with {"status": "abort", "fault": "<fault>"}; aborting without an active fault fails with spurious_abort.
  2. Inner-loop diagnosis (`episode.fork`): Inspect result.grade.violations, result.grade.metrics, and result.log (poses_at, FORCES) to find the earliest causal failure tick, then replay deterministically on the exact seeded T0-T4 plant to tick and hand off to a candidate command or callable:
python
   forked = episode.fork(result, task, tick, candidate_cmd_or_fn)

Confirm fixes on full-episode sweeps after fork debugging.

  1. Paired CRN hill-climbing: Set GAUNTLET_CRN_SECRET (>= 16 bytes) for HMAC-locked evaluation and compare baseline (v0) vs candidate (v1) on identical --task-py, --robot, --tiers=T0,T1,T2,T3,T4, --episodes, and --seed:
bash
   GAUNTLET_CRN_SECRET="$KEY" python -m robosim_gauntlet.cli \
     --task-py=../sim2real-gauntlet/examples/playground_tasks.py --robot=yam \
     --compare-controller="$V0_CMD" --controller="$V1_CMD" \
     --tiers=T0,T1,T2,T3,T4 --episodes=4 --seed=0 --output-json=/tmp/crn.json

Or compare saved runs via --baseline-json=v0.json --candidate-json=v1.json. Verdicts (scorecard.compare): INVALID (fix key/seed/scene_sha256 mismatch), WORSE (revert; safety regressions veto task gains), INDISTINGUISHABLE (keep simpler controller or increase --episodes to reach >= 6 discordant pairs), BETTER (promote only after passing critic_rubric.md, then run --mode=real --mcap-log-dir=...).

同じリポジトリから

関連する Skills

すべての Skills
google-deepmind
コミュニティ

blender-creation

Creates physically calibrated, articulated, PBR-textured 3D MuJoCo assets (MJCF model.xml, .mjz, mjSpec) from scratch using headless Blender (bpy, bmesh, blendermeshutils, assetbuildutils), convex collision synthesis, 10-scenario auditing, and Studio preview. Use only when asked to create, model, or build a 3D MuJoCo object from a URL, image, or text description, or when blender-creation is named. Don't use for non-MuJoCo Blender/Cycles/CAD modeling.

1時間の導入数
2
GitHub Stars
80
更新日
11月17日
google-deepmind
コミュニティ

fr3-duo-control

Franka Duo robot only (Franka Research 3 / FR3 Duo dual-arm cell with Robotiq 2F-85 grippers, ZED Mini and RealSense D405 cameras). Brings up, homes, and controls it via pylibfranka on hardware or in MuJoCo sim. Use only when the user names Franka Duo, FR3 Duo, or fr3duo.

1時間の導入数
2
GitHub Stars
80
更新日
11月17日
google-deepmind
コミュニティ

i2rt-yam-bringup

Checks and installs the I2RT YAM SDK (i2rt), configures Linux socketcan or macOS gsusb CAN bringup for I2RT YAM robot arms, and provides a wave smoke test (wave.py). Use only before bringing up or programming a physical or simulated I2RT YAM robot arm, or when asked to run a wave gesture on an I2RT YAM robot. Don't use for non-YAM CAN bus debugging (e.g. automotive OBD-II or industrial PLC CAN networks) or non-robotics tasks.

1時間の導入数
2
GitHub Stars
80
更新日
11月17日
google-deepmind
コミュニティ

i2rt-yam-ik

Solves 6-DOF inverse kinematics for the I2RT YAM and YAM Pro robot arm using Mink + DAQP on combined YAM + gripper MJCF. Use only when computing joint targets or local Cartesian moves on an I2RT YAM or YAM Pro robot arm. Don't use for non-YAM inverse kinematics, general quadratic programming (QP) solvers, or non-robotics tasks.

1時間の導入数
2
GitHub Stars
80
更新日
11月17日