google-deepmind/robotics-antigravity-plugin

mujoco-playground-rl

Authors, inspects in MuJoCo Studio, audits, trains, and deploys GPU-accelerated RL tasks on MuJoCo Playground and MuJoCo Menagerie 3D robots using PyTorch/Warp reaf-sim flatscene (DeviceLayer / TaskLayer) and RSL-RL.

View source

How to use

  1. Copy the installation command and review the source package before installing it in your own agent environment.
  2. Check the source document for required tools, credentials and supported hosts.
  3. Describe your input and desired deliverable; review the output against the source material.
npx skills add https://github.com/google-deepmind/robotics-antigravity-plugin --skill mujoco-playground-rl

Before you start

  • Installation does not include a model subscription, API allowance or permission to access your files.
  • Repository stars and installation counts describe activity, not an independent quality or security test.
Original skill document

Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.

MuJoCo Playground RL Task Authoring & Training (`rl_task_check.py`)

Create, interactively verify in MuJoCo Studio, audit, train (Impl.WARP + RSL-RL), and deploy new RL tasks on any mujoco_menagerie robot (floating-base quadrupeds/humanoids or fixed-base arms) in mujoco_playground:

  1. Workspace & Accelerator Setup:
  • Clone and sync: git clone https://github.com/google-deepmind/mujoco_playground.git && cd mujoco_playground && make sync (or make sync-cu12 on Linux CUDA 12 drivers < 580, or make sync-cpu on macOS for authoring, Studio inspection, CPU smoke tests, and checkpoint playback; pass --default-index https://pypi.org/simple to uv and --no-sync to uv run).
  1. Robot Scene, MuJoCo Studio Inspection, & `DeviceLayer` (`mujoco_playground/robots/<robot>/sim/{device,scene}.py`):
  • Load spec = asset.get("github:google-deepmind/mujoco_menagerie/<dir>/<file>.xml").spec(). Name unnamed geoms (f"{body.name}_geom_{i}"), ensure any freejoint is named ("root"), add a "terrain" plane geom on worldbody if absent, set spec.option.timestep = 0.004 (control_timestep = 0.02, decimation = 5) and mjINT_IMPLICITFAST, and size cfg.sim.nconmax / cfg.sim.njmax (tolerance = 1e-8, ls_tolerance = 0.01, keeping default solver iteration caps).
  • Replace torque <motor>s with PD position servos (for a in list(spec.actuators): spec.delete(a), add_actuator(name=j, target=j) with trntype=mjTRN_JOINT, gaintype=mjGAIN_FIXED, biastype=mjBIAS_AFFINE, gainprm[0]=kp, biasprm[1]=-kp, biasprm[2]=-kd, ctrllimited=False, forcelimited=True, forcerange=[-effort, effort]; e.g. kp=25.0, kd=0.5, armature=0.01 for quadrupeds), lower high passive torque-model joint damping/frictionloss (e.g. damping=0.1, frictionloss=0.0), filter self-collisions via contype/conaffinity bitmasks, and add an "init_state" standing keyframe.
  • **Interactive Studio Verification (mujoco-studio-sidecar +

mujoco-studio): Load mujoco-studio-sidecar and mujoco-studio** to inspect and hot-reload the robot MjSpec (publishing the spec from SimPlugin.on_load_model in <workspace_dir>/sim_plugin.py, checking standing keyframe stability, PD gains, contacts, and custom ViewerPlugins) in [MuJoCo Studio](sidecar://robotics/mujoco_studio/) by launching <mujoco-studio-sidecar>/scripts/studiod.py --workspace_dir=<dir> with <playground_dir>/.venv/bin/python (IsDaemon: true, checking /_studio/status after edits) before batch training.

  • Wire reaf.simulation.torch (reaf_sim) devices via flat_scene.Component (never pass name=, nworld=, or xpu_device=):
  • reaf_sim.ActuatorDevice(targets=list(JOINT_NAMES), measurements=["joint_pos", "joint_vel"]) ("robot" -> "robot/ctrl", "robot/joint_pos", "robot/joint_vel").
  • Floating base: reaf_sim.FreeDevice(body_name="base", joint_name="root", measurements=["xpos", "xquat", "joint_vel"]) ("base" -> "base/xpos", "base/xquat" [w,x,y,z], and "base/joint_vel" where [:, :3] is world linear velocity -> reaf_sim.math.quat_rotate_inverse(quat, vel[:, :3]), while [:, 3:6] is already body-frame angular velocity).
  • Contacts: flat_scene.Component(reaf_sim.ContactSensorDevice, fields=("found", "force"), history_length=decimation).configure_spec(objtype="geom", objname=..., reftype="geom", refname="terrain") on feet and body/limb geoms.
  1. Sync-Free `TaskLayer` (`mujoco_playground/tasks/<category>/<task>/{task,mdp}.py`):
  • Actions: command_feature_producers={"action_history": flat_scene.Component(reaf_sim.ActionHistory, key="action", feature_shape=(num_joints,), history_length=1)} (f["action"] is a_t, f["action_history"][:, 0] is a_{t-1}) and command_processors={"action_scale": flat_scene.Component(reaf_sim.TransformActions, input_key="action", output_key="robot/ctrl", scale=0.25, offset=DEFAULT_JOINT_POS)}.
  • Stateful Producers & Observations: Subclass reaf_sim.FeatureProducer implementing name, produced_features_spec (per-world shapes without nworld), required_features_keys (consume "time" and step timers by dt = (t - self._last_time).clamp(min=0.0) so partial resets are idempotent), produce_features, and reset; wrap reward/critic-only producers with .mark_privileged() so make_deployment_config prunes them.
  • Sync-Free GPU Rules: Never call .item(), .cpu(), .numpy(), .new_tensor(), bool(), .any(), or .all() inside step/obs_*/reward_*; build gravity via g = torch.zeros((n, 3), device=q.device, dtype=q.dtype); g[:, 2] = -1.0 and cache CPU flat_scene.Obs tensor kwargs per (device, dtype). Configure ObsGroups ("policy" noisy, "critic" clean) with .configure(output_shape=(D,)) on every flat_scene.Obs (actor: body ang_vel, projected_gravity, command, joint_pos - default_pos, joint_vel, action; critic adds body lin_vel).
  • Rewards, Terminations, & Reset: Combine exp(-||v_xy - cmd_xy||^2 / 0.25) (1.5), exp(-(w_z - cmd_wz)^2 / 0.25) (0.75), v_z^2 (-2.0), ||w_xy||^2 (-0.05), ||g_proj_xy||^2 (-2.5), ||a_t - a_{t-1}||^2 (-0.01), ||qvel||^2 (-1e-4), stand-still ||q - q_def||_1 * (||cmd|| < 0.05) (-0.5), and undesired contacts (-1.0); FunctionalRewardProvider multiplies terms by control_dt, so divide touchdown impulses by control_dt. Terminate on falls (g_proj[:, 2] > -0.3 or z < min_z) via torch.where(..., TerminationResult.TERMINATE, TerminationResult.DO_NOT_TERMINATE), and reset standing qpos/qvel in EnvironmentReset.on_reset via torch.where(env_mask.unsqueeze(1), ...).
  1. Register, Validate, Train, Preview in Studio, & Deploy:
  • Add an RslRlOnPolicyRunnerCfg dataclass in _src/rl/rsl_rl/env_rl_config.py (obs_groups={"actor": ["policy"], "critic": ["critic"]}, vecenv_wrapper_cfg=RslRlVecEnvWrappercfg(is_finite_horizon=False), nworld=4096, num_steps_per_env=24, actor/critic.hidden_dims=(256, 128, 128), obs_normalization=True) and register an uncalled env_cfg_fn in _src/rl/env_registry.py.
  • Audit & smoke-test (CPU + torch.cuda.set_sync_debug_mode("error") Impl.WARP + make_deployment_config): uv run --no-sync python <skill_dir>/scripts/rl_task_check.py --repo_root=. --task_dir=mujoco_playground/tasks/<category>/<task> --env_id=<EnvId>.
  • Train & preview in MuJoCo Studio: WP_CACHE_PATH=/tmp/wp_cache uv run --no-sync python -m mujoco_playground.rl.rsl_rl.train --env=<EnvId> --wp_kernel_cache_dir=/tmp/wp_cache --logdir=/tmp/rslrl_logs/<run>. Preview checkpoints interactively in MuJoCo Studio via uv run --no-sync python -m mujoco_playground.rl.rsl_rl.play --env=<EnvId> --resume_path=... --gfx=web --hot_reload=plugins (or --gfx="" for native desktop Studio, authoring custom ViewerPlugins per mujoco-studio), or deploy with flat_scene.make_deployment_config(cfg, observation_groups=("policy",)) + wrapper.TorchDeviceLayerWrapper.
from this repository

More skills

All skills
google-deepmind
Community

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.

1h installs
2
GitHub stars
80
Updated
Nov 17
google-deepmind
Community

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.

1h installs
2
GitHub stars
80
Updated
Nov 17
google-deepmind
Community

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.

1h installs
2
GitHub stars
80
Updated
Nov 17
google-deepmind
Community

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.

1h installs
2
GitHub stars
80
Updated
Nov 17