Robot foundation models
Learning representations and policies that connect vision, language, and action. We investigate how models can transfer skills across tasks and environments.
FLAIR · Research
Our goal is general-purpose physical intelligence: robots that learn from experience, predict the consequences of their actions, and solve tasks that unfold over many steps.
How can robots move beyond individual skills to solve unfamiliar, extended tasks?
We investigate the models and learning algorithms needed to connect perception with physical action. Our current research brings together robot foundation models, world models, reinforcement learning, and agentic robotics.
Learning representations and policies that connect vision, language, and action. We investigate how models can transfer skills across tasks and environments.
Learning generative models of how the physical world evolves. We study physical prediction and spatial reasoning as tools for planning and learning from interaction.
Improving generative policies through reward and interaction. We investigate how to adapt pretrained models while making effective use of experience.
Bringing memory, reasoning, exploration, and planning into robotic systems. We study how robots can combine reusable skills and adapt their plans over long horizons.
Our work on compositional planning studies how learned skills and local generative models can be combined into coherent long-horizon behavior. GSC and CDGS provide two approaches to this problem.
Selected contributions
Physical AI & robotics
Guided search within compositional diffusion connects local models into coherent long-horizon plans.
ICLR 2026
Explore project →Physical AI & robotics
Composing skill-level diffusion models generates long-horizon manipulation plans for unseen task combinations.
CoRL 2023
Explore project →How can generative models learn from feedback and sample efficiently?
We develop algorithms for diffusion and flow models, from reward-based adaptation to sampling in continuous and discrete spaces. These methods connect our work in generative AI and robotics, while addressing questions about generation and inference in their own right.
RVM studies reinforcement learning through velocity matching. DEIS uses exponential integrators to reduce the number of steps needed to sample from diffusion models.
Selected contributions
Generative models & learning
Reward-based velocity matching adapts diffusion models through updates to their velocity fields.
Preprint · 2026
Explore project →Generative models & learning
Exponential integrators exploit the structure of diffusion dynamics to reduce sampling steps.
ICLR 2023
Explore project →Further work includes DMPO for reasoning in diffusion language models and the design of reinforcement learning for diffusion models.
How can we generate, steer, and reason about probability distributions?
We study stochastic control, optimal transport, Schrödinger bridges, sampling, and optimization. Our goal is to understand the mathematical structure of learning and decision-making, and use that structure to design algorithms with provable properties.
This research spans sampling complexity, the control of distributions, and uncertainty in dynamical systems. It supports our Physical AI program and advances theory and methods across control, inference, and machine learning.
Selected contributions
Sampling & mathematical foundations
Approximate rejection sampling improves the dimension dependence of proximal sampling across a broad class of target distributions.
COLT 2023
Explore project →Control & mathematical foundations
A survey connecting stochastic control, optimal transport, Schrödinger bridges, and Sinkhorn algorithms.
SIAM Review · 2021
Explore project →Explore PIS and the connection between sampling and stochastic control, probabilistic reachability analysis, and adjoint Schrödinger bridge sampling.
Find papers, implementations, and opportunities to work with FLAIR.