A curated selection of research shaping physical AI
Stanford, Michigan and Microsoft Research pretrain a force-aware robot foundation model on human tactile data plus 100 hours of robot demonstrations. That robot data is the T-Rex corpus, collected on a Dexmate Vega with Sharpa Wave hands.
Dexmate and Carnegie Mellon introduce MeMoREval, a benchmark for deciding which kind of memory an embodied agent should keep, with evaluations across four simulated environments and a demonstration on a physical Vega 1 Pro.
UIUC and NVIDIA grow a small set of human-object motion captures into a much larger library of motions a robot can execute, then deploy them on real hardware, including a Vega 1 Pro moving a chair.
UIUC, UC Berkeley and Northwestern teach a video world model to predict touch as well as sight. On six contact-rich tasks with a Dexmate Vega and tactile fingertips, it averaged 70.6 against 38.0 for the strongest baseline.
Purdue and Samsung SDS Research America build a real-to-sim pipeline for evaluating vision-language-action policies, then check how well the simulation predicts the real robot. Every experiment, simulated and physical, runs on a Dexmate Vega.
Google DeepMind's on-device robotics model reaches 75.6% task success on a Dexmate robot, up from 33.3% for the previous version.
NVIDIA's CHORD learns two-handed dexterous manipulation from human demonstrations by using the contact forces people apply as guidance. The resulting policies run on a Dexmate Vega with two Sharpa hands.
Northwestern, Peking University and partners build a single simulation runtime where benchmarks, automatic data collection and agent interfaces share one execution loop. A Vega 1 Pro with Sharpa Wave hands is one of its ready-to-run mobile manipulator embodiments.
UC Berkeley and NVIDIA pre-train a vision encoder on 3D hand keypoints from 32,041 hours of human video. On a Dexmate Vega with Sharpa Wave hands it beats the strongest baseline by more than 30 points on real-world dexterous tasks.
USC, Toyota Research Institute and Stanford teach a Dexmate Vega and a Unitree G1 to work as a pair, learning from people who teleoperate both robots at once across four two-robot tasks.
UC Berkeley, NVIDIA and Stanford train a tactile-reactive policy that adjusts its grip as it feels contact. It runs on a Dexmate Vega with Sharpa Wave hands, and the 100-hour dataset collected on that robot is public.
A tactile foundation policy from Tsinghua, Sharpa and partners, pretrained on about 3,000 hours of touch data across 21 sensors. One of its three real-world test rigs is a Dexmate Vega with Sharpa Wave hands, a robot the model had never seen in pretraining.
