DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation
DexTacWAM adds fingertip touch to a video-based robot world model, and the reported gains are large on contact-heavy tasks.
The paper says the system models each fingertip separately, compresses those tactile signals with finger and pose awareness, then feeds them into a video diffusion world model. On six dexterous manipulation tasks using a 22-DoF bimanual platform, it reports a 70.6 average score versus 38.0 for the strongest baseline. The authors say the key gain comes from predicting contact evolution as part of the world state, not just using touch as an extra condition. They also report faster training and inference from the tactile compressor, while keeping most pre-fusion contact recall.
ArXiv · AI/CL/LG's note
The paper says the system models each fingertip separately, compresses those tactile signals with finger and pose awareness, then feeds them into a video diffusion world model. On six dexterous manipulation tasks using a 22-DoF bimanual platform, it reports a 70.6 average score versus 38.0 for the strongest baseline. The authors say the key gain comes from predicting contact evolution as part of the world state, not just using touch as an extra condition. They also report faster training and inference from the tactile compressor, while keeping most pre-fusion contact recall.
ArXiv · AI/CL/LG's note
score 5