MIT & Harvard Researchers Unveil T-WLA, Cutting Finger-Joint MPJPE by 30%

ceciletamura · x · 2026-11-08

Researchers from MIT CSAIL and Harvard (Kaichen Zhou, Yuzhen Chen, Antonio Torralba, Paul Pu Liang, et al.) released T-WLA, a world-language-action architecture addressing the latency gap where contact conditions change faster than visual policies can re-plan.

Key ideas:

Each visual update fuses the current image, task instruction, and retrieved history into a shared visual memory that conditions base-action generation;
The same cached memory feeds a lightweight tactile editor that predicts residual corrections to unexecuted actions from touch, without rerunning the visual branch;
Future-image and future-pressure prediction provide training-only supervision aligning memory and tactile features with action dynamics.

Results: 30% reduction in finger-joint MPJPE vs. the strongest baseline on EgoTouch/Tachin hand-motion prediction, and 50% average success rate across four real-robot dexterous manipulation tasks. Code and videos are public; the authors will discuss the work in a Clubhouse deep dive on Nov 8.

Original post →

MIT & Harvard Researchers Unveil T-WLA, Cutting Finger-Joint MPJPE by 30%

More from Embodied