Simon Fraser University @ ICML 2025

AI has seen tremendous advances in discrete and symbolic reasoning. And yet, reliable AI in the physical world - from robotics, to autonomous driving, to accurate weather prediction to fully automated smart grid - still pose major challenges.

Beyond pragmatic challenges like the availability of data and the costs of operating on physical hardware, this talk will argue that there are unique and fundamental challenges when getting machine learning methods to work on physical systems. Using robotic imitation learning as a didactic example, we will demonstrate mathematical results that reveal error propagation through dynamic environments can lead to exponentially greater data requirements than what we might expect in discrete domains.

On the positive side, we show how certain remedies, like the hitherto-mysterious practice of “action-chunking,” partially surmount these difficulties,  and that a mixture of “imperfect” and “expert” data is strictly better than expert data alone.  Yet, despite these solutions, we will conclude by describing some preliminary work suggesting that there are still major empirical limitations of the state of art methods in continuous control, and will outline some of the most promising directions to overcoming them.

Relevant Work:

“The Pitfalls of Imitation Learning when Actions are Continuous, ” Max Simchowitz, Daniel Pfrommer, Ali Jadbabaie, COLT 2025. ArXiv: https://arxiv.org/abs/2503.09722

The Pitfalls of Imitation Learning when Actions are Continuous;” Thomas Zhang,  Daniel Pfrommer, Nikolai Matni, Max Simchowitz. ArXiv: https://arxiv.org/abs/2503.09722

Rapid improvements over the past few years in computer vision have enabled high performing geometric state estimation on moving camera systems in day-to-day environments. Furthermore, recent substantial improvements in language understanding and vision-language grounding have enabled rapid advancements in semantic scene understanding. In this presentation, I will demonstrate how we can build visual representations from these foundational vision-language models to enable new robotic capabilities in navigation, manipulation, and mobile manipulation. I will also discuss new robotics research directions opened up by these advancements in vision-language understanding and alternative directions to building foundation models for robotics.

Humans have a strong intuitive understanding of the physical world. Through observations and interactions with the environment, we build a mental model that predicts how the world would change if we applied a specific action (i.e., intuitive physics). My research draws on insights from humans and develops model-based reinforcement learning (RL) agents that learn from their interactions and build predictive models that generalize widely across a range of objects made with different materials. The core idea behind my work is to introduce novel representations and incorporate structural priors (e.g., particle- and graph-based neural dynamics models) into learning systems to better model the dynamics of diverse deformable objects. I will discuss how such structures can make model-based planning algorithms more effective and help robots accomplish complicated manipulation tasks (e.g., manipulating an object pile, shaping deformable foam into a target configuration, and making a dumpling from the dough using various tools). Furthermore, I will demonstrate our recent progress in constructing neural and physics-informed digital twins that jointly capture appearance, geometry, and dynamics, with significant implications for data generation and evaluation in the development of robot learning systems.

Curating high-quality data is key for training language models. In this talk, I will present DataComp-LM (DCLM), a large collaboration to build a fully open state-of-the-art pre-training dataset for language models. We will go over each step in the data curation pipeline and highlight the most important parts. For training 7B parameter language models, our DCLM-Baseline training set performs similar to Llama 3 8B on an average of 53 natural language understanding tasks while being trained with 6x less compute.