Live intelligence

Latest AI updates

3 updates
Models

DeepMind unveils Gemini Robotics 2.5 and ASIMOV-jailbreak safety benchmark

Google DeepMind announced Gemini Robotics 2.5, an updated vision-language-action model that generates robotic actions, paired with Gemini Embodied Reasoning, which uses live video to plan tasks from user instructions. DeepMind says the 2.5 release is notably better at high-dexterity tasks, evaluated across more than 200 activities such as serving food with tongs and using a pipette. The company also showed Apptronik's Apollo 2 humanoid tying a bow with five-fingered hands, a task involving deformable objects that is hard for robots. DeepMind additionally released an open safety benchmark called ASIMOV-jailbreak and said some Gemini robots have done actual work in Mercedes-Benz factories.

Google DeepMindGemini RoboticsEmbodied AI
Research

DexDICE reconciles safety and performance in offline imitation learning

DexDICE introduces dual-expert offline imitation learning, using safety-compliant but potentially suboptimal demonstrations alongside high-return demonstrations that may violate safety guidelines. The method combines DICE-style offline learning with safe support constraints rather than conventional distribution constraints. The authors prove that the ideal constrained solution has zero unsafe-region visitation with high probability. They evaluate DexDICE on a re-curated DSRL benchmark and a real-world mobile-robot navigation task trained from human teleoperation, reporting improved performance and robustness to data scarcity and quality degradation.

offline imitation learningrobot safetymobile robot navigation
Research Editor selected

VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression

A new arXiv study examines whether vision-language-action policies refrain from acting when instructions are missing, meaningless, infeasible, or refer to the wrong task. Across six checkpoints in LIBERO-Goal simulation, an empty instruction still caused movement in every episode, and five checkpoints often completed a task the instruction had not specified. Three suppression approaches were tested: an instruction prefix failed across all six checkpoints, an action-chunk gate held the robot on non-task inputs in five of six, and fine-tuning prevented task completion on non-task inputs but did not reliably keep the robot still or handle infeasible instructions.

Why it matters: It provides concrete evidence that current VLA policies can start and execute tasks without valid instructions, an important safety issue for embodied AI deployment.

vision-language-action modelsrobot safetyinstruction following