Particle.news

Microsoft Research Debuts Rho-alpha, a Language-to-Action Model for Two-Handed Robots

Early access trials will test touch-aware control plus human-in-the-loop learning ahead of broader Foundry availability.

Overview

  • Rho-alpha, derived from Microsoft’s Phi series, maps natural-language instructions to precise control for bimanual manipulation.
  • The model extends VLA approaches by incorporating tactile sensing, with work underway to support force signals during operation.
  • Training data combines physical demonstrations with synthetic trajectories produced via NVIDIA Isaac Sim on Azure alongside large-scale visual QA data.
  • Microsoft released demos on the BusyBox benchmark and on plug insertion and toolbox tasks, including a sequence corrected in real time by a human operator.
  • The system is being evaluated on dual-arm platforms and humanoid robots, with a Research Early Access Program open now and a technical description coming in the months ahead.