Intelligence, grounded in reality.

Embodied intelligence concerns how an agent uses perception and interaction to accomplish goals in its environment. Our research investigates how visual observations and human instructions can be translated into purposeful behavior, including navigation and manipulation. This requires representations that connect objects, spatial relationships, and task goals with the actions available to a particular body. Policies must integrate perception, memory, and feedback as observations change or execution diverges from the intended outcome.

A central challenge is generalization across tasks, environments, and embodiments under limited supervision. Our agenda examines learning from demonstrations, multimodal pretraining, and reinforcement learning as complementary routes to transferable skills. We study the relationship between task planning and motor control under partial observability and uncertainty, emphasizing evaluation through task completion, adaptation to unfamiliar conditions, and recovery from execution errors.

World models provide a computational framework for representing how an environment evolves and how actions influence its future. Our research investigates latent dynamics and generative video models that integrate observations over time, infer hidden state, and predict possible continuations conditioned on actions or goals. A key question is how to preserve information relevant to interaction, including object persistence, motion, and the relationship between an agent and its surroundings, despite incomplete observations.

We study how learned simulation can support planning and policy learning through imagined trajectories. This requires careful treatment of uncertainty, accumulated prediction error, and changes beyond the training distribution. Our agenda evaluates temporal consistency, responsiveness to actions, and the accuracy of task-relevant predictions alongside visual quality. The broader objective is to establish when predictive models provide reliable guidance for decisions in unfamiliar situations.

We investigate how images and videos can be transformed into structured representations of three-dimensional scenes and their evolution over time. Our interests include 3D and 4D reconstruction, neural rendering, and generative scene modeling. We study neural fields and explicit representations, including 3D Gaussians, with emphasis on geometric consistency, persistent object identity, and the relationship between appearance and underlying scene structure. These representations should support visual synthesis and queries about spatial relationships.

Physical understanding further requires models of properties and interactions that appearance alone does not determine. We examine how scene representations can incorporate material behavior, contact, and motion by combining learned models with geometric and physical priors. A central question is how uncertainty in reconstructed geometry or inferred physical parameters affects predictions and downstream decisions under incomplete observations.

Multimodal intelligence concerns how complementary sources of information can support a coherent understanding of the world. Our agenda connects vision, language, audio, and spatial representations, with particular emphasis on visual instruction following and reasoning over images and extended video sequences. Central questions include how to preserve fine visual detail, align events across modalities, resolve ambiguous references, and maintain consistent interpretations as observations accumulate.

We further investigate multimodal agents that select external tools, gather additional observations, and revise their responses in light of new information. Linking perception, memory, and task decomposition supports the study of reasoning across multiple steps and modalities. We emphasize compositional generalization, calibrated uncertainty, and mechanisms for detecting and correcting errors, evaluating whether conclusions are supported by observable evidence rather than plausible language alone.