Gemini-4 Argon reaches 81.9% on the RoboVista evaluation, beating GPT-6 Astra and Claude Fable 5.1 with the lowest evaluation cost. It improves across every robotics domain compared with Gemini 3 Flash.
Key points
Score: 81.9% on RoboVista robotics understanding.
Cost: lowest evaluation cost among compared models.
Sources
Contact-UAN – 20% lower torque, 43% lower force error
Contact-UAN adds a residual correction to humanoid actuator commands, achieving 20% lower running torque and 43% lower force-tracking error using only 6.5 min of walking data. It learns the actuator residual from onboard proprioception without torque sensors or motion capture.
Key points
Torque reduction: 20% lower running torque.
Force error reduction: 43% lower force-tracking error.
Sources
Mirror Rust agent core adds python TUI and HTTP API
Mirror is a Rust agent core with a python TUI using an RPC. It can also run headless behind an HTTP API. Every session is a single append-only log.
Key points
Rust agent core with python TUI using RPC
Headless operation available via HTTP API
Sessions stored as single append-only logs
Sources
Grok Bot new improvements announced for weekend
Grok Bot introduces quality, polish, and speed improvements, and chat cards get a cleaner layout. The changes are announced for the weekend.
Key points
Improvements: quality, polish, and speed enhancements to Grok Bot
UI: chat cards have a new design with a cleaner layout
