Locomotion

Humanoid Standup

From the ground up. Evolving neural controllers for a humanoid tasked with standing.

Recordings

292

Archived weight sets

Training clips

232

Ensemble evaluations

Learning progress

run 110888,010

run 138163,711

run 147444,343

run 151662,379

run 1737662,248

run 18504152,763

run 1852462,254

run 1907661,951

run 196413,566

run 3100189,667

run 466862,118

Every logged run as its own mini-curve, best reward per generation; the number shown is that run's final best.

Behavior-map coverage for the best run: the share of the novelty descriptor grid reached as evolution proceeds. 100% = the entire map explored.

Fitness across generations

Each dot represents a training clip: x = filename generation, y = filename fitness (integer-truncated mean training reward). This scatter plot pools recordings without session grouping or per-session traces; dots may overlap and do not form a continuous learning curve.

Each point is one recorded generation from one run; filename fitness is a mean training value, not the return of the clip itself. Points across runs are separate sessions pooled together — the plot is a scatter, not one continuous learning curve.

Featured recording

videos_reward_184010_run_290_2BzH_seed_3_3100_20250516_225024_536501-episode-0.mp4

filename fitness 184,010 · episode 0 · generation 290

Clips by kind

training

232

ensemble

episode