Trustworthy Embodied Foundation Models:
Safety by Design versus Safety by Practice

There is a growing claim that generalist robot policies are becoming more robust. But what does robustness actually mean in this new era? In this talk, I argue that we are overlooking an important notion of “outcome robustness”: even for the same robot action, outcomes can vary depending on the physical world (e.g., dropping an open bag on a table may or may not spill its contents). But, for generalist robots that operate on high-dimensional perceptual inputs, reasoning about this rich space of outcomes requires predicting future observations, which is extremely challenging to model and optimize. To address this, I will present StressDream, an inference-time optimization algorithm that steers diffusion-based video world models toward high-impact yet plausible outcomes by optimizing over the initial noise. In manipulation and autonomous driving with state-of-the-art world models, I will discuss how StressDream enables robust policy evaluation and improvement.

Reliable deployment of robotic foundation models (FMs) requires mechanisms for recovery when learned policies leave the regimes covered by their data. In this talk, I will argue that massively parallel model-based control can provide a common backbone for both safety by design and safety by practice: generating safety-critical data before deployment, and enabling fast intervention during deployment. I will first discuss our work on GPU-parallelized robust MPC, where physics simulation and parallel reachability analysis are used to synthesize controllers and safety certificates for high-dimensional robotic systems, from humanoids to deformable-object manipulation. These robust controllers can be used to produce safe recovery demonstrations for FM finetuning. I will then discuss our recent work showing that classical linear control can be surprisingly effective inside foundation models themselves: by modeling local activation dynamics, we can steer the internal states of world-action models at deployment time to improve robustness under out-of-distribution perturbations. Finally, I will close with perspectives and future opportunities.

Random sampling methods for robot control has demonstrated immense potential in expanding the capabilities of robotic manipulation and locomotion. In this talk, I will show case studies on how this same random sampling can be used to instill robustness in robot behavioral control for manipulation and effectively counteract unobservable environmental uncertainty. With the use of variational inference methods, I show how this robustness can be harnessed to improve the responsiveness of robots in real-time, enabling highly dynamic manipulation with limited sensor-feedback. Last, I overview recent findings on sampling-based hybrid mode composition that draws from hybrid systems theory to extend the capabilities on whole-body robotic control through fast sequencing of diverse control modes.

Generative policies, such as diffusion policies and VLAs, have demonstrated impressive capabilities in solving complex, long-horizon tasks. Yet deploying them in the real world, where robots must continually adapt to novel situations without compromising safety, remains a major challenge. In this talk, I argue for addressing safety and adaptation through a unifying lens: uncertainty awareness. I first show how safety can be enforced on generative policies at deployment time without sacrificing task performance. I then discuss estimating uncertainty in these policies and how it can be leveraged for runtime failure prediction and collecting the most informative data. Finally, I present recent work toward generalist policies that can actively and continually improve.