Frontier signal stream

Skild AI introduced S1, described on its own site as its flagship robotic foundation model "built from the ground up as an in-context learner": shown a single video demonstration of a task, it can attempt previously unseen tasks over roughly 10-minute horizons with no task-specific post-training. Skild frames the advance by analogy to language models — arguing robotics has been "stuck in the BERT era", where each new task still requires collecting hours of data and fine-tuning a specialist policy, and that in-context learning is the transition that turned capable models into general-purpose tools. The stated capability set is: unseen tasks, 10-minute horizons, one video prompt, no post-training.

1X announced 25 degree-of-freedom, tendon-driven hands for the NEO humanoid, which it frames as 'an API to the physical world' and claims approaches human-level dexterity, strength, safety and reliability. Hands are the long-standing bottleneck for general-purpose home manipulation, so this is a component milestone on the critical path rather than a product launch.

1X established a World Model Lab in 2026 to scale humanoid intelligence, naming Sam Sinha as Head of World Models and founding researcher. It signals 1X investing in learned world models as the route to general autonomy for NEO, alongside its existing Redwood AI generalist model.

Boston Dynamics published a technical blog authored by Alberto Rodriguez, Director of Robot Behavior for Atlas, with research engineers Shane Rozen-Levy and Vinay Kamidi, explaining for the first time how Atlas learned to lift and carry unfamiliar objects without prior knowledge of their mass or centre of gravity. Training used GPU-parallel reinforcement-learning simulation across varied weights, friction and grip conditions; the published demonstration uses a 50-lb mini-fridge while internal testing exceeded 110 lb - above the training range - with the behaviour transferring to hardware without retraining (zero-shot sim-to-real).

RLWRLD has announced RLDX-1, a new dexterity-first foundation model designed for complex real-world industrial tasks using high-DoF robotic hands. The model, code, and checkpoints have been made available.

Sanctuary AI published a demonstration of its proprietary hydraulic hand autonomously reorienting a lettered cube to a specified goal orientation ten consecutive times without dropping it, performed entirely at the fingertips without palm support. The company states the manipulation policy transferred zero-shot from simulation to the physical hand. Fingertip-only reorientation is the capability class that underpins precise insertion and tool use, which is the practical bottleneck for industrial deployment of dexterous robots.