RoboChess GymnasiumRoboChess Gymnasium

My honours research at Wits, supervised by Dr S. James and Prof. B. Rosman. The goal: a Gymnasium-compatible reinforcement learning environment supporting a five-task manipulation suite on a simulated UR5e robot arm.

The core design decision was modular reward and observation interfaces, so a new task plugs in without touching the training loop. That constraint is what made the rest of the work tractable.

Running experiments by hand does not scale to a 2×2×2 factorial design at five seeds each. I built a config-driven harness with automated checkpointing, logging, and cluster job submission, which took experiment setup from per-run configuration to a single command across more than 200 training runs.

Results are only worth as much as their error bars. The evaluation pipeline does bootstrap confidence intervals, permutation testing, and Holm–Bonferroni correction, with unit tests over the statistical routines so every reported number holds up under supervisor review.