Exploring Exploration with Foundation Agents | Danny Sawyer, Research Scientist at Google DeepMind
Danny Sawyer is a Research Scientist at Google DeepMind, investigating agentic exploration in text and embodied environments. He worked on the SIMA 2 project, which introduced a multi-domain instructable agent for open-world video games. Previously, he built biologically inspired vision systems for robots deployed in production at Vicarious. He holds a PhD in Bioengineering from Caltech. In this Frontier Research Club talk, Danny presents Exploring Exploration with Foundation Agents in Interactive Environments, new work on how foundation agents explore when they need to actively gather information instead of receiving everything up front. Paper: https://openreview.net/forum?id=wOrkU... The talk focuses on a core question in agentic AI: can foundation agents explore systematically, or do they mostly rely on undirected trial and error? Danny breaks exploration into three measurable capabilities: efficient information gathering, meta-learning, and strategy adaptation. The work tests these capabilities in controlled zero-shot settings, comparing frontier foundation agents against known optimal policies. The presentation uses two environments. Feature World tests whether agents can efficiently gather information in text-based and 3D settings. Alchemy tests whether agents can learn latent causal rules across trials, adapt when those rules change, and use previous experience to improve future decisions. A key finding is that foundation agents can perform well on simpler information-gathering tasks, but struggle more when exploration requires integrating evidence over time. Simple summarization helps models distill what they have learned, improving meta-learning and adaptation in multi-trial environments. Topics include: • Exploration in foundation agents • Interactive agent evaluation • Zero-shot agent behavior • Efficient information gathering • Meta-learning • Strategy adaptation • Feature World • Text-based environments • 3D embodied environments • Alchemy environment • Latent causal rules • Multi-trial learning • Summarization and reflection • Strategy shifts after rule changes • Comparing agents to optimal policies • Foundation agents in dynamic environments Presented at Frontier Research Club by Danny Sawyer. Recorded on March 18, 2026 at Berkeley Skydeck. Frontier Research Club is a curated forum for rigorous discussion on how AI is reshaping the scientific research process. We convene researchers, computational scientists, and research engineers to examine concrete work across literature synthesis, hypothesis generation, experimental design, simulation, analysis, safety, and reproducibility. Upcoming events: https://luma.com/frontiersyndicate Subscribe for more research talks, technical discussions, and frontier AI presentations.

VLM Judges for Autonomous Driving | Jose M. Alvarez, Director of AI Research at NVIDIA

Why Aliens Would NEVER Invade Africa

When millions of AI agents meet

Art of Scaling Reinforcement Learning Compute for LLMs | Bonnie Li, AI Researcher at Google DeepMind

Long Horizon Reasoning | Sumeet Motwani, PhD Student at Oxford

Moondream Segmentation: From Words to Masks | Ethan Reid, Research Scientist at Moondream

Sergey Levine (UC Berkeley): Robot Foundation Models

Everything I Learned Training Frontier Small Models — Maxime Labonne, Liquid AI

Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan

AI, Machine Learning, Deep Learning and Generative AI Explained

Small Batch Size Training for Language Models | Sanae Lotfi, Research Scientist at Meta FAIR

Yann LeCun: World Models: Enabling the next AI revolution

AlphaFold - The Most Useful Thing AI Has Ever Done

What to teach when AI writes the code | Rainer Stropek | TEDxLinz

She Asks if I Know Coldplay and This Singer Shocks The Street

TERMS-Bench: Evaluating LLMs in Semi-Verifiable Domains | Erica Zhang, Stanford PhD

Young Men in Expensive Cars

From Child Prodigy to Winning Fields Medal, Nobel of Math

Scalable Generative Models in Computer Vision

