Roboharness

Approval/evidence harness for unattended robot code changes.

Roboharness is not just a screenshot collector.

The core wedge is:

long unattended agent run -> one proof pack -> short human review

The proving ground starts with the deterministic MuJoCo grasp loop, but the same proof surface also works for humanoid runs across multiple frameworks. From a repo checkout, one command gets back a compiled contract, metric-backed alarms, a phase manifest, an approval report, and an HTML proof surface that tells you what changed and what to do next.

This README preview uses the kept review angles from the same G1 humanoid run: Meshcat front-to-back on the left and MuJoCo top-down on the right. Each frame keeps its phase name visible so you can compare humanoid behavior across frameworks without opening the full report first. To regenerate the same proof surface locally from the committed bundle, run python examples/g1_cross_framework_report.py.

Choose Your Start

Package-First Integration

Use this when you are adding roboharness to an existing codebase. The published wheel installs the library and the roboharness CLI, not the repo's examples/ directory.

pip install roboharness
roboharness --help

The fastest honest package path is the zero-change Gymnasium wrapper shown in Gymnasium Wrapper (Zero-Change Integration). If you want to evaluate the maintained MuJoCo approval wedge itself, use a repo checkout.

Repo Demo: 10-Minute MuJoCo Wedge

This path exercises the shipped MuJoCo approval wedge from this repository.

git clone https://github.com/MiaoDX/roboharness.git
cd roboharness
python -m pip install -e ".[demo]"
python examples/mujoco_grasp.py --report

For headless Linux or CI:

MUJOCO_GL=egl python examples/mujoco_grasp.py --report
# or
MUJOCO_GL=osmesa python examples/mujoco_grasp.py --report

What you get back:

• contract.json — compiled regression contract for this wedge run
• autonomous_report.json — canonical metrics and baseline comparison
• alarms.json — evaluator-backed hard failures
• phase_manifest.json — first failing phase, selected views, rerun hint
• approval_report.json — surfaced vs suppressed case decision for review
• report.html — first-screen proof, not a folder hunt

How to read it:

1. Open report.html.
2. Read the Run Decision banner first.
3. Review only surfaced cases against the old baseline.
4. Use phase_manifest.json and the rerun hint if you need to iterate again.

Baseline rule:

• Regression mode keeps the old baseline authoritative.
• No new baseline is blessed automatically.

Why This Exists

If Claude Code or Codex spends hours refactoring a robot behavior, the hard part is not generating more files. It is getting back one compact proof surface that answers:

• what failed
• where it failed first
• what the current evidence looks like next to the blessed baseline
• whether anything actually needs human review

That is the job of the MuJoCo wedge today.

Installation Matrix

pip install roboharness                  # core (numpy only)
pip install roboharness[demo]            # MuJoCo, Meshcat, Gymnasium, Rerun, Pillow
pip install roboharness[demo,wbc]        # + whole-body control (Pinocchio, Pink)
pip install roboharness[lerobot]         # LeRobot evaluation path
pip install roboharness[dev]             # test/lint/type deps

Progressive Disclosure

• Package-first: wire the wrapper or Harness API into your existing codebase
• Repo demo: from a clone of this repo, run python examples/mujoco_grasp.py --report
• Preset-first: pass --contract-preset mujoco_regression_v1 or --contract-preset mujoco_migration_guarded_v1
• Prompt-assisted: pass --contract-prompt "treat this as migration mode and require manual blessing" to select one of the reviewed presets without opening JSON
• Advanced: pass --contract-json /path/to/contract.json to validate a pre-authored metric-only contract before the wedge starts

If a contract cannot be grounded safely, the run stops before execution and emits a user-facing error envelope with problem, cause, fix, docs_url, and next_action.

Proof Surface

The first screen is meant to be actionable without replay:

• Run Decision tells you whether the run is clean, reviewable, or degraded
• Approval Queue shows changed or ambiguous cases only
• Current vs Baseline shows the first manifest-selected proof pair
• Temporal Evidence appears for ambiguous still-image cases as a checkpoint-ordered strip
• Hard Metric Results shows the evaluator-backed failures
• Evidence images support click-to-zoom for quick inspection without dropping into the gallery
• Phase Timeline and the deeper checkpoint gallery stay available below the fold

pre_grasp	contact	grasp	lift
Gripper above cube	Lowered onto cube	Fingers closed	Cube lifted

View Interactive Reports

• MuJoCo grasp: https://miaodx.com/roboharness/grasp/
• G1 WBC reach: https://miaodx.com/roboharness/g1-reach/
• G1 locomotion: https://miaodx.com/roboharness/g1-loco/
• Native LeRobot GR00T: https://miaodx.com/roboharness/g1-native-groot/
• Native LeRobot SONIC: https://miaodx.com/roboharness/g1-native-sonic/
• SONIC planner: https://miaodx.com/roboharness/sonic-planner/
• SONIC tracking: https://miaodx.com/roboharness/sonic/

Other Demos

These are real integrations and proof surfaces, but they are not the front-door wedge:

Demo	Description	Report	Run
MuJoCo Grasp	Scripted grasp with Meshcat 3D, paired baseline proof, approval report	Live	python examples/mujoco_grasp.py --report
G1 Cross-Framework Proof	Committed Meshcat vs MuJoCo paired-evidence report for one G1 bundle	repo-only	python examples/g1_cross_framework_report.py
G1 WBC Reach	Whole-body IK reaching (Pinocchio + Pink)	Live	python examples/g1_wbc_reach.py --report
G1 Locomotion	GR00T RL stand→walk→stop, HuggingFace model	Live	python examples/lerobot_g1.py --report
G1 Native LeRobot (GR00T)	Official make_env() factory + GR00T Balance + Walk	Live	python examples/lerobot_g1_native.py --controller groot --report
G1 Native LeRobot (SONIC)	Official make_env() factory + SONIC planner	Live	python examples/lerobot_g1_native.py --controller sonic --report
SONIC Planner	Standalone GEAR-SONIC planner demo on G1	Live	python examples/sonic_locomotion.py --report
SONIC Motion Tracking	Real encoder+decoder tracking demo on G1	Live	python examples/sonic_tracking.py --report

Showcase Repository

The showcase repo is for external proof that roboharness works as a pip-installed dependency in real projects:

• LeRobot Evaluation — visual regression testing for robot policies
• GR00T WBC — whole-body control integration

Each showcase is self-contained, runs with ./run.sh, and supports smoke mode for fast CI validation.

pre_grasp	contact	grasp	lift
Gripper above cube	Lowered onto cube	Fingers closed	Cube lifted

G1 Humanoid WBC Reach

pip install roboharness[demo,wbc]
python examples/g1_wbc_reach.py --report

Whole-body control (WBC) for the Unitree G1 humanoid using Pinocchio + Pink differential-IK for upper-body reaching while maintaining lower-body balance. The controller solves inverse kinematics for both arms simultaneously, letting the robot reach arbitrary 3D targets without falling over.

stand	reach_left	reach_both	retract

LeRobot G1 Locomotion

pip install roboharness[demo]
python examples/lerobot_g1.py --report

Integrates the real Unitree G1 43-DOF model from HuggingFace with GR00T WBC locomotion policies (Balance + Walk). The example downloads the model and ONNX policies automatically, runs the G1 through stand → walk → stop phases, and captures multi-camera checkpoints via RobotHarnessWrapper.

Native LeRobot Integration

pip install torch --index-url https://download.pytorch.org/whl/cpu  # CPU-only
pip install roboharness[demo] lerobot

MUJOCO_GL=egl python examples/lerobot_g1_native.py --controller groot --report
MUJOCO_GL=egl python examples/lerobot_g1_native.py --controller sonic --report

Uses LeRobot's official make_env("lerobot/unitree-g1-mujoco") factory for standardized env creation. The published native demo reports are split by controller: one report for GR00T and one for SONIC. DDS-ready for sim-to-real transfer when hardware is available. See #83 for details.

LeRobot Evaluation in CI

pip install roboharness[lerobot]

# Evaluate a real LeRobot checkpoint with visual checkpoints + JSON report
python examples/lerobot_eval_harness.py \
  --checkpoint-path /path/to/lerobot/checkpoint \
  --repo-id lerobot/unitree-g1-mujoco \
  --n-episodes 5 \
  --checkpoint-steps 10 50 100 \
  --assert-threshold \
  --min-success-rate 0.8

Produces:

• episode_000/step_0010/default_rgb.png — checkpoint screenshots
• lerobot_eval_report.json — structured per-episode stats
• CI exit code 1 when thresholds are not met

SONIC Planner

pip install roboharness[demo]
MUJOCO_GL=egl python examples/sonic_locomotion.py --report --assert-success

Standalone NVIDIA GEAR-SONIC planner demo on the real Unitree G1 MuJoCo model. This path uses planner_sonic.onnx only: velocity commands go in, full-body pose trajectories come out, and the example uses a lightweight virtual torso harness for stable visual debugging. This is the same standalone planner path published at /sonic-planner/.

from roboharness.robots.unitree_g1 import SonicLocomotionController, SonicMode

ctrl = SonicLocomotionController()
action = ctrl.compute(
    command={"velocity": [0.3, 0.0, 0.0], "mode": SonicMode.WALK},
    state={"qpos": qpos, "qvel": qvel},
)

For a planner demo wired through LeRobot's official make_env() stack, see G1 Native LeRobot (SONIC) above. The planner path and the encoder+decoder tracking path are different inference stacks with different ONNX contracts; see docs/sonic-inference-stacks.md for the exact split, validation policy, and joint-order conventions.

SONIC Motion Tracking

pip install roboharness[demo]
MUJOCO_GL=egl python examples/sonic_tracking.py --report --assert-success

Real encoder+decoder tracking demo on the Unitree G1. This path uses model_encoder.onnx + model_decoder.onnx directly, replays a motion clip via set_tracking_clip(...), and records checkpoint metrics for torso height, tracking-frame progress, and joint-tracking error. This is the same path published at /sonic/.

from roboharness.robots.unitree_g1 import MotionClipLoader, SonicLocomotionController

ctrl = SonicLocomotionController()
clip = MotionClipLoader.load("path/to/dance_clip/")
ctrl.set_tracking_clip(clip)

action = ctrl.compute(
    command={"tracking": True},
    state={"qpos": qpos, "qvel": qvel},
)

Models (planner_sonic.onnx, model_encoder.onnx, model_decoder.onnx) are downloaded from HuggingFace (nvidia/GEAR-SONIC) on first use. Requires pip install roboharness[demo]. See docs/sonic-inference-stacks.md for the exact split between planner and tracking, plus the validation policy and joint-order conventions. See #86 (Phase 1) and #92 (Phase 2).

Gymnasium Wrapper (Zero-Change Integration)

import gymnasium as gym
from roboharness.wrappers import RobotHarnessWrapper

env = gym.make("CartPole-v1", render_mode="rgb_array")
env = RobotHarnessWrapper(env,
    checkpoints=[{"name": "early", "step": 10}, {"name": "mid", "step": 50}],
    output_dir="./harness_output",
)

obs, info = env.reset()
for _ in range(200):
    obs, reward, terminated, truncated, info = env.step(env.action_space.sample())
    if "checkpoint" in info:
        print(f"Checkpoint '{info['checkpoint']['name']}' captured!")

Core Harness API

from roboharness import Harness
from roboharness.backends.mujoco_meshcat import MuJoCoMeshcatBackend

backend = MuJoCoMeshcatBackend(model_path="robot.xml", cameras=["front", "side"])
harness = Harness(backend, output_dir="./output", task_name="pick_and_place")

harness.add_checkpoint("pre_grasp", cameras=["front", "side"])
harness.add_checkpoint("lift", cameras=["front", "side"])
harness.reset()
result = harness.run_to_next_checkpoint(actions)
# result.views → multi-view screenshots, result.state → joint angles, poses

Supported Simulators

Simulator	Status	Integration
MuJoCo + Meshcat	✅ Implemented	Native backend adapter
LeRobot (G1 MuJoCo)	✅ Implemented	Gymnasium Wrapper + Controllers
LeRobot Native (make_env)	✅ Implemented	make_env() + VectorEnvAdapter
Isaac Lab	✅ Implemented	Gymnasium Wrapper (GPU required for E2E)
ManiSkill	✅ Implemented	Gymnasium Wrapper
LocoMuJoCo / MuJoCo Playground / unitree_rl_gym	📋 Roadmap	Various

Design Principles

• Harness only does "pause → capture → resume" — agent logic stays in your code
• Gymnasium Wrapper for zero-change integration — works with Isaac Lab, ManiSkill, etc.
• SimulatorBackend protocol — implement a few methods, plug in any simulator
• Agent-consumable output — PNG + JSON files that any coding agent can read

See ARCHITECTURE.md, CONTRIBUTING.md, CHANGELOG.md, docs/development-workflow.md, and docs/context.en.md for architecture, contributor workflow, release notes, and project background.

Related Work

Roboharness builds on ideas from several research efforts in AI-driven robot evaluation and code-as-policy:

• FAEA — LLM agents as embodied manipulation controllers without demonstrations or fine-tuning (Tsui et al., 2026)
• CaP-X — Benchmark framework for coding agents that program robot manipulation tasks (Fu et al., 2026)
• StepEval — VLM-based subgoal evaluation for scoring intermediate robot manipulation steps (ElMallah et al., 2025)
• SOLE-R1 — Video-language reasoning as the sole reward signal for on-robot RL (Schroeder et al., 2026)
• AOR — Multimodal coding agents that iteratively rewrite control code from visual observations (Kumar, 2026)

Citing

If you use Roboharness in academic work, please cite it using the metadata in CITATION.cff or the "Cite this repository" button on GitHub.

Contributing

Contributions welcome — including from AI coding agents! See CONTRIBUTING.md.

License

MIT