Lilly Tech Systems
Humanoid Robots Best Practices
Practical guidance for building, testing, and deploying humanoid robots safely and effectively.
Development Pipeline
Design and Simulation
Start in simulation (Isaac Sim, MuJoCo). Validate kinematics, dynamics, and control algorithms before touching hardware. Use URDF/MJCF models of your robot.
Policy Training
Train locomotion and manipulation policies in massively parallel simulation. Use domain randomization for robustness.
Hardware-in-the-Loop Testing
Test with the real robot tethered to a safety gantry. Validate sensor readings, actuator responses, and latency.
Controlled Deployment
Deploy in a controlled environment with safety barriers. Gradually expand the operational envelope as confidence grows.
Field Operation
Deploy in the target environment with monitoring, remote e-stop, and rollback capabilities.
Safety Checklist
- Hardware e-stop: Physical button accessible to all nearby humans. Must cut actuator power immediately.
- Software safety layer: Independent safety controller that monitors joint limits, velocities, and forces, running at a higher priority than the AI policy.
- Collision monitoring: Real-time obstacle detection and automatic deceleration or stop.
- Failsafe posture: A safe default pose the robot assumes when errors occur (e.g., crouching with arms tucked).
- Logging and replay: Record all sensor data and commands for post-incident analysis.
Sim-to-Real Best Practices
| Technique | What to Randomize | Typical Range |
|---|---|---|
| Mass randomization | Link masses | ±15-30% of nominal |
| Friction | Ground and object friction coefficients | 0.3 to 1.5 |
| Actuator delay | Control loop latency | 0 to 30ms |
| Sensor noise | IMU bias, joint encoder noise | Realistic noise profiles |
| Visual randomization | Lighting, textures, camera position | Wide variation |
Testing Strategies
- Unit tests: Test individual components (perception, planning, control) in isolation with recorded data
- Integration tests: Run the full pipeline in simulation with scripted scenarios
- Stress tests: Push the robot to edge cases — unexpected obstacles, adversarial perturbations, sensor failures
- Long-duration tests: Run for hours to catch memory leaks, thermal issues, and drift
- Human-in-the-loop tests: Have untrained users interact with the robot to find usability issues
Frequently Asked Questions
NVIDIA Isaac Sim offers the best GPU-accelerated parallel simulation for RL training. MuJoCo is excellent for research and fast prototyping. PyBullet is free and good for learning. Choose based on your GPU resources and whether you need photorealistic rendering.
Domain randomization is the primary technique: randomize physics, visuals, and dynamics during training so the policy is robust to real-world variations. System identification (measuring real-world parameters and matching simulation) also helps. Finally, a small amount of real-world fine-tuning can close remaining gaps.
Dexterous manipulation, long-horizon task planning, robustness to unexpected situations, battery life, and cost. While locomotion has made great progress, reliable autonomous operation in truly unstructured environments remains an open problem.