Robots on their Best BEHAVIOR – Fireside Chat With Fei-Fei Li & Jim Fan

Join Stanford’s Fei-Fei Li and NVIDIA’s Jim Fan for an in-depth discussion of BEHAVIOR, a large-scale benchmark and challenge for advancing embodied AI. The session will outline the scientific vision and research motivations behind BEHAVIOR, highlight how it links perception, reasoning, and action in real-world household settings, and explain the design of the BEHAVIOR Challenge. It will also cover evaluation methodologies, the differences between standard and privileged information tracks, and the role of simulation in advancing robotics research. This is an opportunity to understand how BEHAVIOR unifies efforts across academia and industry to build robust, human-centered intelligent agents​​. Key Moments: 00:18 — Welcome & Agenda | BEHAVIOR benchmark overview and goals (perception ↔ reasoning ↔ action) 01:37 — Guest Intros | Dr. Fei-Fei Li (Stanford HAI, World Labs) · Dr. Jim Fan (NVIDIA, GR00T) 03:30 — ImageNet Lessons | North-star tasks & the big-data bet that reshaped AI 06:14 — Introducing BEHAVIOR 1K | Large-scale simulation benchmark & challenge for embodied AI 09:12 — Task Selection | From American Time Use Survey + 1,000-person study to top human-wanted robot tasks 15:18 — Simulation Stack | OmniGibson on Omniverse/Isaac: ~50 interactive scenes, ~10k objects, rigid + deformable + fluids 18:54 — Why Scale & Sim Matter | Safer data, combinatorial coverage; closing sim-to-real with modern CV & generative 3D 24:02 — Why Robotics Benchmarking Is Hard | Action taxonomy, context, and need for a unifying north star 29:40 — Fireside Q&A Starts | Ethics, values, and practical simulation questions 36:15 — Call to Action | Join the BEHAVIOR Challenge; GTC DC announcements & resources Q&A Highlights: 29:46 — How are human values/ethics encoded (e.g., in GR00T or BEHAVIOR-trained systems)? A: BEHAVIOR’s tasks were human-centered (survey-driven, assistive focus). Models inherit priors from the benchmark and data design; values must be infused across the pipeline, not in one step. 32:52 — For startups building RL environments, what simulation aspect matters most? A: It depends on the task: navigation vs. manipulation, rigid vs. deformable/fluids demand different fidelity and throughput trade-offs; optimize sim for your target behaviors and gaps. Resources: - Project site (BEHAVIOR-1K) — overview, tasks, challenge, docs. (https://behavior.stanford.edu/) - Paper (arXiv) — BEHAVIOR-1K: human-centered, 1,000 activities; OmniGibson sim. (https://arxiv.org/abs/2403.09227) - Code/datasets — BEHAVIOR-1K repo (https://github.com/StanfordVL/BEHAVIOR-1K) - Install guide — hardware reqs + OmniGibson/Isaac dependencies (https://behavior.stanford.edu/getting_started/installation.html) Got questions, post them on our Discord Thread: https://discord.com/channels/827959428476174346/1423891601389916171 Discord Invite: https://discord.com/invite/nvidiaomniverse 📆 Check out the full calendar for all of our upcoming events → https://nvda.ws/3JqaWnA ---------------------------------------------------------------------------- ⬇️Get Started → https://nvda.ws/4cZAZO1 ​ 👀Explore OpenUSD → https://nvda.ws/3CeozBQ ​ 👥Join the Community → https://nvda.ws/3ZMfc6e