Tag: Dexterous manipulation

Video as Policy: Markov Robotics Controls a Hand With LTX-2.5

Markov Robotics uses the LTX-2.5 video model as a manipulation policy: the robot generates footage of the completed task and executes the motion it implies. A 1:192 compression ratio makes it computable, but no benchmarks have been published.

GEN-1.5: a robot learns a new task from one demo

Generalist AI unveiled GEN-1.5 — a robot foundation model that learns a new task from a single 3–12 s demo, with no fine-tuning. One-shot yields 59% average success, few-shot reaches 83%.

Unichem Buys Loomia to Enter the Humanoid Skin Market

South Korea's Unichem acquired New York-based Loomia for 8.6 billion won (about USD 6 million) to enter the market for electronic skin and tactile sensing for humanoids.

A robot finger that feels in color: a 100-µm tactile sensor

A team from Queen Mary University of London and Italian universities built a tactile sensor that turns pressure into color via a mechanochromic Bragg reflector — 100 µm resolution with zero computational latency.

Dyna-2: a million hours of human video gives robots zero-shot skill

Dyna Robotics unveiled Dyna-2 — a world-action model trained on over a million hours of human video that the company says demonstrates the first human-to-robot scaling law.

T-Rex: a manipulation model that reacts to touch

A team from Stanford, UC Berkeley and NVIDIA — with Li Fei-Fei among 34 authors — unveiled T-Rex, a vision-language-action model that adds high-frequency touch. It lifts success rates by over 30 points across 12 dexterous tasks.

Tacta Systems debuts a touch-sensing dexterous robot hand

Palo Alto startup Tacta Systems came out of stealth on July 28, 2026 with TactaBot — a 15-joint robotic hand with dense tactile sensing, a data-collection glove and an AI model. The target is precision manufacturing assembly, shipping in early 2027.

RLDX-1: a foundation model for robot hands with memory and force sensing

RLWRLD has released RLDX-1 — a foundation model built from the ground up for high-DoF robot hands. The model integrates context memory, force sensing, and motion processing into a single architecture, achieving up to 90% success rates on tasks requiring physical signals and task history.

NVIDIA i PSYONIC mostem nad „luką danych”: Nowy standard zręczności robotów