ACT (Action Chunking with Transformers)

Pose-Aware Modeling to Mitigate Pose-Related Artifacts in Tactile Gloves

Tactile gloves digitize contact and force during hand-object interactions, but pose-related artifacts degrade data quality. This work proposes pose-aware modeling to mitigate such artifacts, improving tactile sensing reliability for robotics applications.

Real2Sim2Real for Vision-Language-Action Manipulation: An AMD ROCm-Based Pipeline

A pipeline for vision-language-action (VLA) manipulation using Real2Sim2Real transfer, built on AMD ROCm for efficient simulation-to-real deployment.

OAT Tokenization: A New Interface for Visuomotor Policies That Balances Compression and Decodability

Ordered Action Tokenization (OAT) introduces a learned tokenizer that maps continuous robot actions into ordered discrete tokens, achieving high compression, total decodability, and an ordered token space. Tested across 60+ tasks and multiple policy backbones, OAT enables anytime inference tradeoffs and outperforms existing discretization methods. For builders, this means more flexible, efficient visuomotor policies without sacrificing performance.

StARS: Socially Appropriate Robot Actions via a Recommender System-Driven Approach

StARS reformulates socially appropriate action generation as a preference modeling problem using recommender systems, capturing inter-subject variability in HRI.