UC Berkeley and NVIDIA Unveil Agentic Robotics as NASA Recruits

Breakthroughs in autonomous physical systems took center stage this week as UC Berkeley engineering researchers presented "agentic robotics" alongside NVIDIA's open simulation and world action models at the Conference on Robot Learning (CoRL) in Austin, Texas, while NASA opened applications for its 25th astronaut class to staff upcoming Artemis lunar missions.

UC Berkeley's Ken Goldberg Unveils Graph-as-Policy and Agentic Robotics

UC Berkeley Professor of Engineering Ken Goldberg, recently inducted into the National Academy of Engineering, teamed with Spencer Huang, director of robotics engineering at NVIDIA, to introduce "agentic robotics"—an approach that bridges classical model-based control and data-heavy imitation learning to overcome what Goldberg previously termed the field's "100,000-year data gap." Comparing the shift to the debut of the World Wide Web in 1993 and ChatGPT in 2022, Goldberg described how teams of AI agents generate, test, and refine interpretable robot-control programs in simulation.

Using a framework called Graph-as-Policy (GaP) together with OpenAI's Astra model and NVIDIA's Isaac simulation libraries, postdoctoral researchers Eric Kaiyuan Chen and Shuangyu Xie supplied a short video of a robot using a sponge to wipe ink off a metal bar. In less than an hour, the AI built a physics simulation modeling the sponge's deformation, revised the control code after an initial trial pushed over the bar, and executed the task on hardware.

Berkeley-based logistics company Ambi Robotics—where co-founder and CTO Jeff Mahler leads warehouse automation—applied the agentic robotics workflow to production parcel-sorting data. Testing demonstrated an increase of more than 15,000 packages sorted annually per robot, and the update is now deployed across 30% of Ambi Robotics' U.S. fleet.

NVIDIA Launches GR00T 2, Isaac Lab 3.0, and VABAR Benchmarks

At CoRL, where two NVIDIA papers—GRAIL and RoboTTT—were named among six Best Paper Award finalists, NVIDIA released GR00T 2, an open frontier world action model built on NVIDIA Cosmos and trained on 100,000 hours of human egocentric data. Out of the box, GR00T 2 completed 40 of 50 evaluated manipulation tasks at least once without task-specific demonstrations, achieved 71% average success across 50 tasks with 1 demonstration per task, and reached 95% success with 5 demonstrations, ranking among the top three models on the RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces benchmarks.

NVIDIA also introduced its Video to Data workflow, the EgoHand 3D hand-motion model (which has logged over 40,000 Hugging Face downloads in two weeks with data from Noitom Robotics), and Isaac Lab 3.0 integrated with the open-source Newton physics engine, delivering up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. Using Success-Guided Sampling developed with the University of Washington across more than a million parallel environments, the AlphaNIST policy learned all 20 precision tasks on the NIST Assembly Task Board with more than 98% training success—compared with 81% for separate per-task policies—using roughly 39x less GPU time. In addition, NVIDIA rolled out Isaac Lab-Arena 0.4 alongside RLWRLD's 15-task, four-scene DexBench benchmark and UC Berkeley's six-task Variational Automation Benchmark for Agentic Robotics (VABAR), with partners including Foxconn, Siemens, Bosch Research, Skild AI, Flexion, FieldAI, Dyna Robotics, Walden Robotics, Dexmate, and MicroAGI.

NASA Opens Applications Through December 3 for 25th Astronaut Class

Expanding robotic and human exploration beyond low-Earth orbit, NASA announced at a Kennedy Space Center air show in Cape Canaveral that applications are open through December 3, 2026, for its 25th astronaut class since the original Mercury Seven in 1959. In its previous recruitment cycle two years ago, more than 8,000 candidates applied and 10 were selected in September 2025, leaving the agency with 32 active astronauts today.

Following April's lunar flyaround by the Artemis II crew, NASA plans to select the new class by late 2027 and begin two years of mandatory training in early 2028—the same year NASA intends to return astronauts to the lunar surface. Artemis III commander Randy Bresnik and crewmate Frank Rubio—who holds the U.S. space endurance record of 371 consecutive days in space, and who noted that flying to the Moon represents about 1% of an astronaut's overall mission support work—will launch into Earth orbit next summer with two other crewmates to practice docking with SpaceX and Blue Origin lunar landers ahead of the Artemis IV south pole landing mission.