UMD's PAIR framework bridges perception and action in VLA models via bridge tokens
Kaixi Feng · hf · 2026-10-16
Researchers at University of Maryland (Kaixi Feng, Guoheng Sun, Ang Li) propose PAIR, a framework addressing the implicit perception-to-action transition in continuous-action vision-language-action models.
Method
- A Masked Action Autoencoder encodes expert action chunks into horizon-aligned Action Latent Tokens during training;
- A Bridge Module extracts task-relevant features from visual-language representations and aligns them with the latent tokens to form Bridge Tokens preserving task info and expert action structure;
- Bridge Tokens are projected into action-token space and injected into initial Action Tokens; at inference the autoencoder is removed and Bridge Tokens come only from observation and instruction.
Results: gains on LIBERO, LIBERO-Plus, and CALVIN ABC→D benchmarks.
More from Embodied
- Tesla Roadster Thruster Teardown: SpaceX COPV Tank Plus Optimus Actuator — XFreeze · 2026-10-16
- Critic Calls Figure's Terminator-Style Robot Demo Publicity Over Substance — FuSheng_0306 · 2026-10-16
- Snap Spectacles + Gemini 4 Argon demo designs LEGO builds from your brick pile — stspanho · 2026-10-16
- Waymo Announces Its Robotaxi Service Is Coming to Toronto — ChrisArchitect · 2026-10-16
- Polymarket odds of Tesla shipping Optimus this year jump 5 points to 12% after flying Roadster reveal — Polymarket · 2026-10-16
- More companies sell robot intelligence than robots themselves, as Dexmate's Vega becomes a research hub — lukas_m_ziegler · 2026-10-16