What happened
The Workshop on Agentic Robotics took place on November 12, 2026, at CoRL 2026 in Austin, Texas. The workshop website names Agentic World Action Models, by UC Berkeley, NVIDIA and the University of Pennsylvania, among the Best Paper Award winners. EmbodiedSWE: Coding Agents for Long-Horizon Dexterous Robotics, from Yale, Princeton, Stanford, CMU, ByteDance Seed and collaborators, is also named a winner. The organizers say the Best Paper Award was sponsored by Waddle Labs and carried a $1,000 prize.
The workshop frames agentic robotics as the composition of reasoning, skills and verification at inference time on top of learned policies. It argues that bringing this agent paradigm into physical robots creates problems not present in digital settings: actions are continuous and can be irreversible, observations are noisy and partial, and decisions must happen in real time. Its stated purpose was to bring together communities that often work separately, including foundation-model and vision-language-action researchers, planners, control and policy-learning researchers, and software-agent researchers.
The organizers identify six open problems. These include the boundary between an agent’s reasoning layer and a robot’s motor policy; how systems might manage real-world self-improvement; where to route feedback so it generalizes; how to use simulation as a source of data and verifiable rewards; whether zero-shot composition requires a factored agent-policy design; and how to benchmark multi-turn systems across simulation and reality. The website poses these as research questions, not as settled findings or capabilities demonstrated by the award announcement.
The event combined invited talks, posters, live demonstrations and audience-driven discussion. The schedule named speakers from Tencent Video AI Center, Carnegie Mellon University, UC Berkeley, Google Gemini Robotics, NVIDIA and Dexmate. The website describes diagnostic tasks spanning tabletop manipulation, mobile manipulation and navigation, intended to expose seams between perception, reasoning, skill selection, control and recovery. That framing places the award within a workshop designed to compare not just model outputs, but operational behavior across a robot system.
The organizers also opened a demo track for systems using on-site platforms. They say a robot, compute and an API for perception and control were ready at the venue, allowing participants to run an agentic system without transporting their own hardware. Submissions required video proof of a system running and a description of hardware and software requirements. The workshop closed with an award ceremony; its organizers describe a community position paper as a later workshop artifact.
What the archive adds
New Runtime covered NVIDIA’s GR00T 2 and Isaac Lab 3.0 at CoRL on November 9, 2026, including robotics workflows and benchmarks. This award is a new event in the same broader robotics research thread, not a new capability announcement for GR00T 2.
Trend signal
This supports agent economics moves to completed work only weakly: the workshop’s questions about recovery and honest benchmarking point toward evaluating what a robot system completes, not just its model calls. But the award page reports no task-completion measurements or cost comparisons, so it does not yet demonstrate that shift.
Why now
The award appeared on the workshop website after the November 12, 2026 event. It is a timely marker of which work a cross-disciplinary robotics workshop chose to recognize, while the organizers’ open-problem framing shows that practical architecture and evaluation standards remain contested. The live-demo format also reflects a field trying to test behavior under interaction rather than relying only on static paper claims.
Connections to other events
The award sits alongside New Runtime’s earlier coverage of GR00T 2 and Isaac Lab 3.0 at CoRL, which described NVIDIA’s robotics model and workflow announcement. Together, the stories touch both deployed research infrastructure and research on how agents should reason, act and recover with physical systems.
Why it matters
For teams building AI systems that interact with the physical world, a good result cannot be defined only by whether a model produces a plausible plan. Robot actions can be irreversible, observations incomplete, and timing constraints real. The workshop’s agenda makes clear that teams need to decide what belongs in a reasoning layer versus a motor policy, where corrective feedback should go, and how to test recovery without rewarding benchmark gaming.
Those choices affect system design before a team selects a model. A language-capable planner may be useful for task decomposition and exception handling, while a lower-level learned policy may need to handle fast motor control. If an evaluation records only eventual success, it can hide unsafe retries, brittle recovery behavior, or excessive human intervention. Builders need tests that make those distinctions visible.
The award itself is a recognition, not evidence that these questions have been solved. Still, the workshop format and its focus on shared diagnostics offer a useful signal: progress in robotics agents will depend on evaluating the complete loop from perception through action and recovery, rather than treating language reasoning or a successful demonstration as sufficient.
Limits and uncertainty
The workshop website names award winners but does not summarize the winning paper’s methods, experiments or results. It therefore does not establish what Agentic World Action Models contributes, how it performs, or whether its approach transfers across robots and tasks. The award is not independent validation of deployment readiness.
The website describes open questions and planned community discussion, but does not report agreed answers, benchmark outcomes, or results from the live demonstrations. It does not state the evaluation criteria used for the Best Paper Award or compare the two winning projects. A future position paper could clarify the workshop’s common definitions and proposed measurements, but no such publication is described as complete.
Reader takeaway
The award is a research signal, not a ready-to-use robotics capability. The durable lesson for builders is the workshop’s focus on the seams: planning versus motor control, feedback versus generalization, and task success versus genuine recovery. Those are the engineering questions to watch as agents move from digital tools into physical systems.