Investigación

learning

Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success

The paper proposes Sensorimotor Alignment (S²), a training-free offline metric that compares a model’s sensory representations with the geometry of physical motion, represented by end-effector trajectories. Across 12 vision-language-action models, S² positively correlates with downstream robot task success. The authors present it as a low-cost way to prioritize models and identify promising vision-language backbones before fine-tuning or costly closed-loop evaluation.

Traducción pendiente · Resumen en inglés

arXivOct 22, 2026

Ilustración de IA

learning

Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success

The paper introduces Sensorimotor Alignment (S²), a training-free offline metric that measures how closely an embodied model’s sensory representations align with the geometry of physical motion. The authors report that S² correlates positively with downstream robot task success across embodied model families and can help prioritize vision-language model backbones before costly fine-tuning into vision-language-action models. It is proposed as a proxy, not a replacement for closed-loop evaluation.

Traducción pendiente · Resumen en inglés

arXiv cs.ROOct 22, 2026

learning

SCOPE: Scene-graph Construction Optimized by Plan Error-correction for Robot Long-term Planning Datasets

SCOPE is an automated framework for generating diverse, executable task-and-motion-planning datasets for robots. It uses planner execution failures to correct logical inconsistencies in scene graphs, aiming to ensure plans are physically feasible. In experiments, data generated by SCOPE improved a fine-tuned downstream TAMP model’s success rate by an absolute 12.33% over a counterpart trained on DELTA-generated data.

Traducción pendiente · Resumen en inglés

arXiv cs.ROOct 22, 2026

learning

Skeleton-Grounded Latent Action Model for Robot World Models

SkeLAM is a skeleton-guided latent action model for generating robot observations conditioned on actions. It separates robot kinematics from residual scene dynamics, uses skeleton supervision and cross-modal reconstruction to learn the kinematic representation, and adapts robot actions to that space. On DROID, it reduces end-effector ADE by 19.5% and LPIPS by 11.1% relative to DreamDojo; the authors also report zero-shot human-to-robot motion transfer in generated rollouts.

Traducción pendiente · Resumen en inglés

arXiv cs.ROOct 22, 2026

learning

Learning Robot Behavior Priors from Non-Robotic Synthetic Data

The paper tests whether a Transformer can learn useful robot-action priors from procedurally generated data based on mathematical abstractions of control, without seeing real robot behavior during pretraining. The model achieves non-trivial zero-shot action prediction on unseen physical robot tasks through in-context learning; specialized control priors outperform generic causal priors. A compute-matched model trained on real robot data performs better only on the datasets it saw during training.

Traducción pendiente · Resumen en inglés

arXiv cs.ROOct 22, 2026

Verified published records only; coverage is not a census. arXiv preprints may not be peer reviewed. Official code is linked only after verification.