Anthropic's 64.9% Dual-Arm Result, Apple's HALO Lift Robotics Shares

Two figures set the day's frame. Anthropic's Fable 5.2 reports 64.9% success on dual-arm manipulation tasks, and Skild AI has open-sourced Steel, a physics solver it says runs up to 13 times faster than PhysX at roughly one-third the GPU memory. One number describes policy capability, the other simulation throughput per unit of compute.

The Anthropic result reached equity markets faster than it reached any lab benchmark: A-share robotics names rallied on Nov 10, and Apple's HALO robot-brain preprint separately lifted Unitree and Qwen-linked stocks. Frontier-lab preprints now function as a pricing input for Chinese embodied AI assets, a transmission channel that did not exist a year ago.

Infrastructure releases cluster around data and simulation economics. Robbyant shipped weights, code and a technical report for LingBot-Touch, a tactile model pretrained on nearly 5,000 hours of real contact data, while Steel lowers the compute floor for physics-heavy training. The binding constraint is migrating from model architecture toward contact-data volume and solver cost per rollout.

Hardware validation stayed narrow and quantitative. Boston Dynamics mounted wheels on Atlas to measure how faithfully a reinforcement-learning whole-body controller tracks commands through real edge actuators, with the emphasis on measurement fidelity. Qiyuan Robot's 10,000th unit came off the line on Nov 9, the first Chinese production milestone at that order of magnitude.

Leaderobot tallied 20 robotics financing deals worth about RMB 6 billion in one week, with allocation spreading beyond humanoids; Qingtianzu and Hangzhou Qiantang New Area state capital formed a RMB 12 million joint venture, and J.P. Morgan Asset Management's China team visited AGIBOT. Two observation points follow: whether preprint-to-price transmission persists, and whether tactile data scale displaces manipulation success rate as the headline benchmark.