G1 Load & Slope Risk Copilot

A learned world model + walking controller stack for the Unitree G1 humanoid. It predicts fall risk from real-time MuJoCo state and now runs a learned PPO walking policy in the loop.

Status: World model is trained and verified (test AUC 0.9971). PPO walking policy v4 is training on the Nebius VM; the 1.5M-step checkpoint walks ~1.5 m on flat ground and handles mild slopes for ~1 s.

---

What it does

1. Simulates Unitree G1 in MuJoCo on slopes with external forces.
2. Predicts per-timestep fall risk with a GRU world model.
3. Controls the robot with either a heuristic, safe-MPC, reference gait, or PPO policy.
4. Serves rollouts through a FastAPI backend with a web-based trajectory viewer.
5. Benchmarks controllers across terrain suites (flat, slope, SF hills).

---

Repo layout

.
├── configs/                    # Training YAML configs
├── data/                       # Generated MuJoCo rollouts (not in git, ~1 GB)
├── frontend/                   # React + FastAPI demo
│   ├── backend/
│   │   ├── app.py              # FastAPI server
│   │   ├── mujoco_rollout.py   # Real MuJoCo episode runner
│   │   ├── ppo_controller.py   # Loads trained PPO policy
│   │   ├── safe_controller.py  # MPC safety layer over heuristic
│   │   ├── risk_model.py       # GRU fall-risk scorer
│   │   └── benchmark.py        # Production benchmark suite
│   └── src/                    # React frontend
├── models/                     # Trained checkpoints (not in git except noted)
├── src/
│   ├── train_world_model.py    # GRU training
│   ├── infer.py / evaluate.py  # Offline inference & metrics
│   ├── mujoco_collector/       # MuJoCo data collection & controllers
│   └── mujoco_rl/              # PPO training env + script
├── DATA_SCHEMA.md              # CSV column docs
├── NEBIUS_SETUP.md             # VM setup notes
└── requirements.txt

---

Quick start

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# 1. Generate / collect MuJoCo data
python -m src.mujoco_collector.collector --episodes 2048 --workers 16

# 2. Train the fall-risk world model
python src/train_world_model.py --config configs/mujoco_valloss.yaml

# 3. Verify independently
python verify_model.py

# 4. Train the PPO walking policy
python src/mujoco_rl/train_ppo.py --output models/g1_ppo_walk_v4 --steps 5000000 --n-envs 16

# 5. Run the backend
python frontend/backend/app.py

---

World model results

Metric	Value
Test AUC	0.9971
Test F1	0.9798
Test accuracy	0.9744
Architecture	2-layer GRU + MLP head, ~220k params
Input	10-timestep sliding window of state + slope + force
Best checkpoint	models/mujoco_valloss/best_model.pt

The backend bundles a copy at frontend/backend/model/mujoco_valloss/best_model.pt.

---

Controllers

Controller	Description	Status
heuristic	Hand-tuned balance + velocity tracking	Falls frequently on hard terrain
safe	Heuristic + 5-candidate MPC safety layer	Reduces risk, still falls on steep slopes
reference	Phase-based sinusoidal gait	Work in progress
ppo	Stable-Baselines3 PPO trained in MuJoCo	Best current policy; v4 training

Switch controllers via the backend rollout endpoint (controller_type) or benchmark:

python frontend/backend/benchmark.py --controllers ppo safe heuristic

---

PPO training

Training happens on the Nebius L40S VM inside tmux session ppo_v4:

ssh hemad@195.242.29.248
tmux attach -t ppo_v4
# or tail the log
tail -f /tmp/ppo_v4.log

Final flat-ground performance (seed 42, speed 0.5 m/s) after 5M steps:

Metric	Value
Duration	2.54 s
Forward distance	3.34 m
Improvement over heuristic	5.6×

On hills the policy degrades gracefully (e.g. 2.2 m on 5°, 1.6 m on 10°), which matches expectations since it was trained on flat ground.

---

Demo

The backend is running on the Nebius VM. Expose it locally with an SSH tunnel:

ssh -L 8000:localhost:8000 hemad@195.242.29.248

Then open the frontend or hit the API at http://localhost:8000.

Backend API

The FastAPI server runs on http://localhost:8000.

Generate a rollout:

curl -X POST "http://localhost:8000/rollout" \
  -H "Content-Type: application/json" \
  -d '{"seconds":8,"incline_deg":0,"friction":1.0,"speed_mps":0.5,"controller":"ppo"}'

Response contains frames with joint angles, root transform, and risk_score per frame.

---

Benchmark suites

frontend/backend/benchmark.py runs 500 episodes across:

• easy_flat
• medium_slope
• hard_slope
• filbert_street (31.5% grade)
• bradford_street (41% grade)

Outputs CSV + JSON summary in results/benchmark/.

---

Notes

• data/*.csv and most models/ directories are gitignored because they are large (100 MB–1 GB). The small backend model copy in frontend/backend/model/ is included.
• The final PPO v4 checkpoint (g1_ppo_final.zip) is bundled at frontend/backend/model/g1_ppo_walk_v4/ so the backend works out of the box.
• Unitree’s official ONNX velocity policy was unstable in this MJCF (trained on IsaacLab USD), so we retrained PPO directly in MuJoCo.

---

Team

• Pratham
• Hema
• Mahek