Robotics research gets more explicit about memory and mechanism

Dual-Anchoring makes long-horizon vision-language navigation more explicit about task state. The core idea is simple: force the model to state which instruction sub-goals are done, and force it to retain a landmark-level memory of where it has been. That supervision is large-scale, with 3.6M progress samples and 937K grounded landmark samples. The reported gains are strong on continuous-environment VLN benchmarks: success rate reaches 65.6 on R2R-CE and 61.7 on RxR-CE, with about +8.7 and +8.8 points over StreamVLN. In this small period, that makes memory and progress tracking the clearest algorithmic result.

MM-Hand focuses on hardware practicality in dexterous manipulation. It is a 21-DOF open-source hand with remote tendon actuation, modular 3D-printed structure, quick tendon connectors, and room for richer sensing in the hand itself. The engineering trade-off is measured, not hidden: a 1 m sheath gives 25 N fingertip force versus about 33 N with a 0.1 m sheath, and the controller still holds steady-state joint error below 0.1° with about 0.2 s delay. The paper also reports that friction matters more than arm-motion disturbance in its tracking tests. That makes this work useful as a research platform paper, not just a concept demo.