UC Berkeley's AdaDexGrasp lets robots grasp like humans
New robotics AI fuses vision and touch to handle fragile objects 3x better
Researchers from UC Berkeley's Department of Electrical Engineering and Computer Sciences have unveiled AdaDexGrasp, a groundbreaking framework for robotic grasping that mimics human-like adaptability by integrating visual and tactile data.
The system introduces a unified visuo-tactile representation that maps tactile signals to specific robotic fingers, enabling contact-aware grasp planning and real-time adjustments after contact. Unlike traditional methods that rely solely on vision, AdaDexGrasp continuously refines its grip using tactile feedback, achieving significantly higher success rates—particularly for delicate or irregularly shaped objects. In experiments, the approach demonstrated robust performance across 100+ diverse objects, including fragile items that previously posed challenges for robotic systems.
The team validated AdaDexGrasp in both simulated environments and real-world robotics setups, showing its potential for applications in logistics, manufacturing, and assistive robotics. The research has been accepted to ECCV 2026, highlighting its significance in advancing dexterous manipulation.
- AdaDexGrasp combines vision and tactile feedback in a single 3D representation for adaptive grasping
- System achieved 3x higher success rates on fragile/irregular objects vs. vision-only methods
- Validated in both simulation and real-world robotics with 100+ diverse test objects
Why It Matters
Enables robots to handle delicate items with human-like precision, unlocking automation in logistics and manufacturing