aselimc Sim-to-Real Transfer
Domain Randomization
- Visual: texture, lighting, camera pose, background
- Dynamics: friction, mass, damping, actuator delay
- Observation noise: Gaussian noise on sensors, dropout
- Action: smoothing, latency simulation (1-3 step delay)
Strategies
| Approach | When | Pros | Cons |
|---|---|---|---|
| Domain randomization | No real data | Simple, scalable | Over-conservative |
| System identification | Some real data | Accurate | Requires instrumentation |
| Fine-tuning in real | After sim training | Best performance | Needs safe exploration |
| Asymmetric actor-critic | Privileged sim info | Uses GT in training only | Complex implementation |
Hardware Abstraction
Design interfaces that work identically in sim and real:
class RobotInterface(ABC):
def get_observation(self) -> dict: ...
def send_action(self, action: np.ndarray) -> None: ...
def reset(self) -> dict: ...
Key Libraries
Isaac Sim/Lab, MuJoCo, Gazebo, PyBullet