Robots enter the era of fine-motor dexterity

Dexterity in production. Google DeepMind’s Gemini Robotics 2.5 platform demonstrated high-precision manipulation by successfully performing over 200 distinct activities, including pipette handling and serving food with tongs (Ars Technica). Apptronik’s Apollo 2 humanoid, powered by this model, achieved the difficult task of tying a bow with five-fingered hands, marking a shift toward handling deformable objects in factory settings. Meanwhile, AGIBOT tested six G2 humanoid units in a 24-hour convenience-store challenge, where the machines autonomously stocked shelves and prepared coffee for customers (The Rundown AI, The Rundown AI).

Scaling physical intelligence. Nvidia’s GR00T 2 model reached a 71% success rate across 50 evaluated tasks using only one demonstration per activity, a figure that climbed to 95% with five demonstrations (AI Weekly). Researchers in Japan developed 3D-printed humanoid legs that replicate human muscles, tendons, and ligaments to improve movement fluidity in dynamic environments. Amazon has established a dedicated Personal Robotics Group, acquiring Fauna Robotics to develop a 3.5-foot-tall humanoid capable of navigating complex residential spaces (TLDR Tech). Finally, a Caltech robot successfully assembled a structure in orbit, demonstrating the capability to build solar-power stations in space autonomously.

• MIT and NYU researchers created the FLEX system, which allows robots to interpret the intent behind instructions to avoid damaging electronics (The Rundown AI).
• Researchers found that tampering with a robot’s camera feed significantly increases the probability of the machine performing dangerous actions in home environments.