VLA Models

AIAI-generated summary NVIDIA introduces Alpamayo 2 Edge, an adaptive reasoning vision-language-action model designed for real-time operation on NVIDIA DRIVE AGX Thor. NVIDIA says the model improves closed-loop driving performance by 48.9% through reinforcement learning in NVIDIA AlpaGym.

AIAI-generated summary Google DeepMind reports that Gemini 4 Argon robotics research capabilities acted as a simulation engine to accelerate the development of Gemini Robotics 2.5. Google DeepMind says Argon generates simulated practice runs from camera scenes and builds digital twins to help robots generalize to new objects.

AIAI-generated summary Google released Gemini 4 Argon as a general-purpose physical-reasoning model intended for robot control, simulation, video labeling, and policy improvement. Google reports that Gemini 4 Argon achieves an 87 percent score on the Embodied Reasoning Question Answering benchmark and is available to subscribers and through the Gemini API.

AIAI-generated summary At the Conference on Robot Learning in Austin, Texas, NVIDIA announced new open robotics technologies and models. NVIDIA says the offerings include GR00T 2 and Isaac Lab 3.0 to help researchers train robots for complex tasks.

AIAI-generated summary The Conference on Robot Learning features keynotes comparing Fei-Fei Li's world models from World Labs with Russ Tedrake's factory-trained behavior models from Walden Robotics. World Labs describes Atlas as a multimodal autoregressive diffusion transformer that generates 3D scenes, while Walden Robotics reports its robots have worked in a Toyota plant.

AIAI-generated summary AGIBOT unveiled its GO-3 Preview robot model on Oct. 29 at its Innovation Day in Shanghai, featuring a system built on a world reasoning action model. AGIBOT says the model ranks first on three benchmarks, though the company supplied those rankings itself and has not yet released the weights or code to the public.

AIAI-generated summary LiberAI announced Liber-0, a world action model pretrained on more than 2.2 million hours of human experience to support zero-shot robot manipulation. LiberAI says the model achieved an overall success rate of 36.0% across 60 tasks evaluated in 600 trials without robot demonstrations collected specifically for those target tasks.