Rho-alpha (ρα)

Microsoft Research · 2026-01(2차)†

VLANew in 2026Not disclosedNot disclosed

The first robotics model derived from the Microsoft Phi family, a 'VLA+' that adds tactile sensing.

Paper

Backbone LLM: Derived from a Phi-family VLM
Vision encoder: —
Action head: Not disclosed (Not disclosed)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): Derived from a Phi-family VLM; visual·proprioceptive·tactile inputs; action representation not disclosed

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed
Data type: Real-robot teleoperation demonstrations + simulation·synthetic (Isaac Sim RL trajectories) + web vision-language (VQA)
Data scale: Not disclosed
Main datasets: Commercial·open-source demonstration datasets (names not disclosed)
Tasks & evaluation: BusyBox benchmark, bimanual UR5e with tactile sensors, humanoid; figures not disclosed
Robot: Bimanual UR5e (tactile sensors), humanoid

Reinforcement learning: Partial (RL-generated data) · Synthetic trajectories generated by RL in Isaac Sim used for co-training. Continual learning from human teleoperation feedback during deployment
Base control: Fixed · No mention of base·locomotion (centered on tactile bimanual UR5e; humanoid mentioned only as an evaluation platform)

Key features: The first robotics model derived from the Microsoft Phi family, a 'VLA+' that adds tactile sensing. Learns continually from human teleoperation feedback during deployment.
Notes: Release date 2026-01-25 is secondary (TechRadar); the primary page has no date, so it could not be confirmed. Research early access (Microsoft Foundry). Force sensing is in development.
Verification: Could not be confirmed