LabEvolver applies training-free experience evolution to physical agents
LabEvolver explores safe, grounded adaptation for embodied agents operating in wet-lab environments without conventional retraining.
Event Type: research publication
Social navigation learns from human-leg sensing
A navigation approach uses human-leg sensing to improve robot behavior around people in shared spaces.
Decentralized JEPA predicts a shared future for robot teams
CS-JEPA lets distributed robots predict a common future collective state from local observations and limited messages.
World Models Receive a Security Threat Model
Researchers map cyber-physical attacks across sensing, world modelling, planning, control and human interaction.
Action-Conditioned World Models Move Into Planning
The work uses action-conditioned predictive models to improve generalizable embodied decision-making.
VLA Failures Become Corrective Training Signals
RedFlow converts failed robot actions into localized corrections for vision-language-action policy improvement.
Robots Are Asked to Reason About Their Own Aging
A framework integrates hardware degradation into autonomous planning, mission execution and maintenance decisions.
Safe Agent Execution Gets Localized Obligations
HALO proposes localized obligations for admitting and supervising heterogeneous agent actions safely.
Temporal Constraints Target More Stable Latent Actions
The research introduces temporally constrained latent-action learning for more coherent embodied control.
A Humanoid Head Gains Personalized Episodic Memory
German researchers evaluated lightweight personalized memory for sustained human-humanoid interaction.
