README.md

---
license: apache-2.0
tags:
  - robotics
  - navigation
  - reinforcement-learning
  - imitation-learning
  - dagger
  - humanoid
  - unitree-g1
  - isaaclab
  - sim-to-real
language:
  - en
pipeline_tag: reinforcement-learning
---

# G1 Navigation Policy (DAgger Distillation)

Vision-based navigation policies for the **Unitree G1 humanoid robot**, trained using teacher-student DAgger distillation.

## Model Description

This repository contains two PyTorch TorchScript models:

| Model | File | Input Dim | Output Dim | Description |
|-------|------|-----------|------------|-------------|
| **Student** | `student_policy.pt` | 82 | 3 | Deployable policy using depth rays from monocular depth estimation |
| **Teacher** | `teacher_policy.pt` | 34 | 3 | Privileged policy with ground-truth obstacle positions |

### Architecture

Both policies use a 3-layer MLP with ELU activations:
- **Student**: [82] → 256 → 128 → 64 → [3]
- **Teacher**: [34] → 256 → 128 → 64 → [3]

### Observation Spaces

**Student (82-dim):**
- Depth rays: 72 dims (±70° FOV, corrupted with noise/dropout)
- Robot velocity (vx, vy, ω): 3 dims
- Goal relative position: 2 dims
- Goal distance & angle: 2 dims
- Previous action: 3 dims

**Teacher (34-dim):**
- Nearest obstacles: 8 × 3 = 24 dims (x, y, distance)
- Robot velocity: 3 dims
- Goal relative + distance/angle: 4 dims
- Previous action: 3 dims

### Action Space

Velocity commands: `[vx, vy, ω]`
- `vx ∈ [-0.6, 1.0]` m/s (forward/backward)
- `vy ∈ [-0.5, 0.5]` m/s (lateral)
- `ω ∈ [-1.57, 1.57]` rad/s (yaw rate)

## Training Details

### Two-Stage Pipeline

1. **Stage 1: Teacher PPO** - Train privileged teacher with ground-truth obstacles using PPO (2000 iterations)
2. **Stage 2: DAgger Distillation** - Distill teacher to student using Dataset Aggregation with 70% → 20% teacher mixing decay

### Key Innovations

- **FOV Randomization**: `fov_keep_ratio ∈ [0.35, 1.0]` prevents sensor overfitting
- **Hardcase Curriculum**: Mine failure trajectories, retrain with 35% hardcase resets
- **Symmetry Augmentation**: 50% mirror transform eliminates left/right bias
- **Runtime Safety Layer**: Distance-based velocity scaling for collision avoidance

### Performance

| Scenario | Success Rate | Collision Rate |
|----------|-------------|----------------|
| Deploy (mild noise) | 75.0% | 25.0% |
| Stress (heavy noise) | 75.2% | 24.8% |
| Wide-FOV Clean | 74.4% | 25.6% |

### Real Robot Validation

| Direction | Target | Final Distance | Result |
|-----------|--------|----------------|--------|
| Forward | 2.0m | 0.26m | ✅ SUCCESS |
| Backward | -1.5m | 0.26m | ✅ SUCCESS |
| Left | 1.5m | 0.31m | ✅ SUCCESS |
| Right | -2.0m | 0.26m | ✅ SUCCESS |
| Diagonal | (1.5, 1.5)m | 0.30m | ✅ SUCCESS |

## Usage

```python
import torch

# Load student policy (for deployment)
student = torch.jit.load("student_policy.pt")
student.eval()

# Prepare observation (82-dim)
obs = torch.zeros(1, 82)  # [depth_rays(72), vel(3), goal_rel(2), goal_dist_angle(2), prev_action(3)]

# Get action
with torch.no_grad():
    action = student(obs)  # [vx, vy, omega]
```

## Training Environment

- **Simulator**: NVIDIA IsaacLab (Isaac Sim 4.5)
- **Arena**: 8m × 8m with 24-32 cylindrical obstacles
- **Control Rate**: 10 Hz policy / 50 Hz physics
- **Robot**: Unitree G1 (capsule proxy in sim, full robot for real deployment)

## Citation

If you use this model, please cite:

```bibtex
@misc{g1-navigation-dagger-2026,
  title={Teacher-Student Distillation via DAgger for Sim-to-Real Navigation on the Unitree G1},
  author={Adjimavo},
  year={2026},
  url={https://huggingface.co/Adjimavo/g1-navigation-dagger}
}
```

## License

Apache 2.0