Coevolved robot communication transfers poorly to 3D: 1 success in 30 seeds
uv-mex · hf · 2026-10-10
A Universidad Veracruzana study tests transferring a co-evolved communication protocol from 2D simulation to a 3D physical environment without retraining.
- Two e-puck-type robots run a GRU network with residual connections on a food-seeking task with social signaling.
- The sensory/motor translation layer needed three corrections, including calibrating a hunger term from a measurable asymmetry in residual weights.
- Transfer was partial and asymmetric: across 30 seeds, one agent reached food once; the other never did.
- Adding explicit directional information to the social channel changed trajectories and helped some cases, but not enough.
Conclusion: successful transfer of emergent communication depends not just on preserving signaling, but on the ecological and navigational conditions under which the protocol evolved.
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
- Transferring co-evolved robot communication from 2D to 3D physics simulation — 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