Applies to v0.1.0.

Domain adaptation

The Sim-to-Real step: closing the gap between synthetic and real camera frames so a policy trained in Nano-sim holds up on the physical car.

The gap

Synthetic frames differ from real ones in lighting, texture, and noise. A policy that overfits to the simulator's look degrades on real footage. This is the core Sim-to-Real problem this project is built around.

Current approach: noise injection

Today, transfer relies on Gaussian noise injection during training: perturbing the simulated observations so the policy stops depending on the simulator's exact appearance. This produced the lane-following agent used on the International Robot Contest simulator track, which transferred to the real vehicle with minimal fine-tuning.

Planned: diffusion-based adaptation

Planned. A diffusion model that translates synthetic frames toward the appearance of real ones is the next step. This section will be expanded once it is implemented.