Unitree Go2 Tier 1 — paid embodied skills in MuJoCo + PyBullet + Webots

Scope: simulator-only submission. No physical robot is involved; the x402 payment gate and the wire contract are exercised end to end in peer-mode Zenoh, and on-chain settlement (Base Sepolia, EIP-3009) is exercised live and provably.

A paid RoboPay action arriving on the tunnel's Zenoh topic starts a Go2 skill episode on the official MuJoCo model (google-deepmind/mujoco_menagerie unitree_go2). Nine skills are available — wave, sit, stand, stop, bow, nod, turn_to_face, hold and navigate_obstacle — each driven by a joint-space PD trajectory controller with gravity compensation, never a recorded motion or built-in demo. The same joint configurations are recomputed in PyBullet (from a kinematic URDF generated deterministically from the same go2.xml) and compared, so the paid action is a real, measured embodiment in both simulators.

The chain, top to bottom:

paid action (x402 / AIP) -> tunnel -> Zenoh "robot/tunnel/action"
-> subscriber -> validate envelope + x402 payment gate -> Go2 skill
-> joint PD on the mujoco_menagerie model -> metrics
-> result on "robot/tunnel/result" (correlated by actionId)

Every successful skill returns the body to the home stance height afterwards (|bodyZ - 0.283| < 0.02), so paid actions can run back to back.

Steering uses a measured calf-gain calibration (STEER_TABLE in go2_control.py): the shared calf gain kc scales calf = -1.8 + kc*off and produces a monotone, straight-line net heading over -21.7°..0° — so a descending course is followed as clean, low-drift segments. A potential-field local planner pulls toward a look-ahead point on the waypoint segment and repels from obstacles. Obstacle contact is detected by the MuJoCo physics engine (contact pairs on obs_* geoms), never a distance estimate. TIMEOUT and COLLISION are proper error results; test_adversarial_nav.py proves both on the real controller path (unreachable goal -> TIMEOUT; blocking obstacle -> COLLISION, 8 simultaneous contact pairs).

Measured on the committed course (simulation/docs/obstacle_nav_report.json, course drawn in obstacle_course_map.svg):

paid action (x402) → tunnel → Zenoh robot/tunnel/action
    → robopay_link.py → validate envelope + payment gate (durable replay)
    → joint-space Go2 controller on mujoco_menagerie
    → metrics → result on robot/tunnel/result (correlated by actionId)
    → settle ONLY on status:success (local ledger; optional Base Sepolia)

• Real action, not a demo — each skill is a joint-space trajectory driven by the controller; physics metrics are measured (paw lift, sit depth, torso pitch/yaw, body height, achieved heading, obstacle contacts), not scripted.
• Payment safety — settle only on status: success; unpaid ⇒ 402 + PAYMENT-REQUIRED, forged/expired receipts ⇒ 402, replay ⇒ 409, tampered paramsHash ⇒ INVALID_PARAMS. Every failure path returns an error result and never settles (test_payment_gate.py, test_result_semantics.py).
• Sim-to-sim — the same joint configurations are recomputed in PyBullet from a kinematic URDF generated from the same go2.xml; foot-tip positions agree to 1 cm tolerance with observed worst-case 0.02 cm (simulation/pybullet/go2_sim2sim_report.json). A Webots R2025a supervisor harness is committed and honestly SKIPs when the runtime is missing (simulation/webots/).
• Durable replay — idempotency keys / txHashes survive a store restart (test_durable_replay.py).
• Live on-chain settlement (Base Sepolia, EIP-3009) — 3 real transferWithAuthorization transactions settled 1.0 USDC each on Base Sepolia (chainId 84532, USDC 0x036CbD53842c5426634e7929541eC2318f3dCF7e), funded entirely from free faucets; no-settle-on-failure proven on-chain (relay nonce unchanged). Evidence: simulation/docs/settlement-proof.json
  • settlement-proof-failure.json.
• Reproducible — clean checkout + bash simulation/verify_go2_tier1.sh, under 30 minutes.

pip install mujoco>=3.1.3 numpy pybullet cryptography eclipse-zenoh eth-account
cd simulation
./setup.sh                      # pinned official menagerie unitree_go2 assets
cd go2
python3 test_go2_control.py     # every skill's physics actually happen
python3 test_payment_gate.py    # 402/409, settle-only-on-success
python3 test_result_semantics.py# success/error semantics, replay, tampering
python3 test_link.py            # paid action → Zenoh → episode → result
python3 test_obstacle_nav.py    # calf-gain steering + potential-field nav
python3 test_adversarial_nav.py # honest TIMEOUT / COLLISION failure matrix
python3 test_durable_replay.py  # replay keys survive a store restart
python3 test_settlement.py      # EIP-3009 offline proof + no-settle-on-failure
cd ../pybullet
python3 test_sim2sim_go2.py     # MuJoCo ⇄ PyBullet agreement (≤0.02 cm)
cd ../webots
bash run_webots_sim2sim.sh      # real Webots R2025a sim-to-sim (or honest SKIP)

One command: bash simulation/verify_go2_tier1.sh runs every acceptance test above and exits nonzero if any fails.

export BASE_SEPOLIA_RPC_URL="https://sepolia.base.org"
export PRIVATE_KEY="0x..."          # payee private key (NEVER commit!)
export PAYEE_ADDRESS="0x..."        # derived from PRIVATE_KEY if not set
python3 test_payment_gate.py        # will attempt Base Sepolia settlement on success

registry/vendors/unitree/go2/unitree.go2.mujoco-pybullet-sim.v1/
    robot.profile.yaml      robot identity + Zenoh runtime
    skills.yaml             the 9 skills, params, limits
    functions.yaml          agent REST contract (/action, 402)
    payment-policy.yaml     x402 pricing + settle-on-success rule
    execution-mapping.yaml  skill → simulator runtime + metrics
    examples/               paid action envelope
    tests/                  skill-contract cases
    docs/                   README + validation report
simulation/
    setup.sh                pinned fetch of official menagerie unitree_go2
    go2/                    Go2 controller, payment gate, Zenoh link, obstacle
                            navigator, durable replay, settlement module + 9
                            test suites
    pybullet/               kinematic URDF (generated from go2.xml) + sim-to-sim
    webots/                 Webots R2025a supervisor harness (honest SKIP
                            without the runtime)
    docs/                   evidence: go2.gif, CI logs, shots, obstacle reports,
                            settlement proofs

• Screen recording: simulation/docs/go2.gif
• Per-skill screenshots: simulation/docs/go2-shots/
• Obstacle course map (real physics trajectory): simulation/docs/obstacle_course_map.svg
• Obstacle navigation report: simulation/docs/obstacle_nav_report.json
• Adversarial failure matrix: simulation/docs/obstacle_adversarial_report.json
• Sim-to-sim report: simulation/pybullet/go2_sim2sim_report.json
• Live settlement proof: simulation/docs/settlement-proof.json + settlement-proof-failure.json
• CI logs: simulation/docs/go2-ci-logs.txt

• Simulator-only profile; the x402 gate mirrors the tunnel's middleware decision semantics in Python (not the compiled Go binary).
• Base Sepolia settlement is env-gated (needs funded payee key + RPC); the committed proofs are from real faucet-funded runs.
• Webots sim-to-sim is best-effort: without the runtime it writes an honest SKIP report and never fakes a pass.
• turn_to_face uses a static-stability hip-abduction shuffle (honest residual error reported). navigate_obstacle uses static obstacles only.

The full reproducible environment (Python deps + pinned Go2 model assets + one-command verification) ships as a public container image:

docker pull ghcr.io/eslam-x/robopay-go2-tier1:latest
docker run --rm ghcr.io/eslam-x/robopay-go2-tier1:latest

The image runs verify_go2_tier1.sh (every acceptance test, exit nonzero on failure) and is rebuilt from each v* tag by .github/workflows/publish-docker-image.yml.

Open source, licensed MIT. New skills, navigation features, and payment-gate hardening are welcome. Start with CONTRIBUTING.md and the good first issue label; every change must keep bash simulation/verify_go2_tier1.sh green. Please read the code of conduct and report vulnerabilities privately per SECURITY.md.

This repository is the original authorship archive of the Unitree Go2 Tier-1 submission to the RoboPay bounty (PR #89). It lives independently of any fork so the work is provably authored by EslaM-X, timestamped by git history, and tagged for release. Every file in registry/ and simulation/ is the authored submission content; see the ROADMAP for the archive's principles.

MIT — © 2026 EslaM-X 🇪🇬