Sensorimotor Alignment: Evaluating Robotic Models Offline
This paper introduces Sensorimotor Alignment (S²) as a cost-effective offline method to evaluate robotic models by measuring alignment between sensory representations and physical motion.
IAEI / RESEARCH & IDEAS
Research briefs on embodied intelligence, robotics and learning through interaction, with links to original sources.
FOLLOW THE RESEARCH
Briefs are automatically prepared from public research abstracts. They do not imply independent verification or association endorsement.
This paper introduces Sensorimotor Alignment (S²) as a cost-effective offline method to evaluate robotic models by measuring alignment between sensory representations and physical motion.
AdaDex is a goal-conditioned RL framework that combines motion tracking and generalization to learn adaptable dexterous skills from short human demonstrations.
EventSync is a hierarchical framework that uses event-driven memory updates to improve history-dependent robotic manipulation by coordinating task reasoning and action execution.
This research explores whether joint sensing alone can enable precise robotic assembly tasks without vision, force, torque, or tactile feedback.
CoDA is a framework that co-evolves simulation data engines and action policies using shared memory to improve robot training and out-of-distribution success.
P³-WAM is a unified model for egocentric dexterous manipulation that integrates 3D perception, action prediction, and goal-directed planning.
LIBERO-PeRM introduces a benchmark for evaluating personalized robotic manipulation, focusing on preference learning in embodied agents.
This paper introduces SafeDriveVLA, a vision-language-action model designed to enhance autonomous driving by predicting navigation compliance modes and using navigation-conditioned world model dreaming to ensure safety.
ChunkAct introduces action chunking to improve efficiency in agentic robotics by reducing the need for repeated LLM reasoning.
Chronos Policy addresses the limitations of fixed-rate observation-action sequences in robot policies by making physical time explicit, enhancing adaptability to varying observation rates and inference latencies.
Splat2Act is an automated pipeline that personalizes a pretrained vision-language-action policy to a specific scene and object set using a handheld phone scan and images of the user's objects.
RESCUE is a counterfactual framework to assess whether alarms from runtime monitors enable useful interventions in robot policies.