Transferring co-evolved robot communication from 2D to 3D physics simulation
uv-mex · hf · 2026-10-10
A companion study examines whether a communication mechanism co-evolved between two robotic agents in a discrete 2D simulator retains its function after transfer to a 3D physics-based simulator.
- The study analyzes effects of episode time budget, social cues, and asymmetry between co-evolved roles.
- Success rate rose roughly linearly with time budgets, with no plateau between 2,000 and 6,000 physics steps—shorter-episode evaluations may underestimate trained controllers.
- In both simulators, the social cue worked mainly as jam-assistance rather than navigation, more pronounced in 2D.
- Eight evolutionary runs showed consistent asymmetry direction with varying magnitude; controlling agent processing order ruled out physics-engine artifacts.
Factors behind the remaining post-transfer performance gap are discussed.
More from Embodied
- Former SpaceX engineers raise $100M for self-driving electric freight trains — chrisgrayson · 2026-10-10
- Kleiner Perkins' Ilya Fushman: the moment for robotics is finally now — Sethwinterroth · 2026-10-10
- Dex-One2Many: Real2Sim2Real Turns One Human Video Into Robots Generalizing Across Configurations — furongh · 2026-10-10
- Coevolved robot communication transfers poorly to 3D: 1 success in 30 seeds — uv-mex · 2026-10-10
- SGS tweak to RL resets lets sim-trained robots mesh gears at 94% zero-shot — abhishekunique7 · 2026-10-10
- Blogger suggests adding CUA data to robotic model pre-training mix — eigenron · 2026-10-10