We propose TESPEC, a self-supervised pretraining framework designed to capture long-term spatio-temporal information from event data. Unlike existing approaches that pretrain feedforward models on short event streams, TESPEC is tailored for recurrent models and leverages long event sequences. Using a masked image modeling strategy and a novel event accumulation method, TESPEC reconstructs pseudo-grayscale videos. Our model achieves state-of-the-art performance in downstream tasks such as object detection, semantic segmentation, and monocular depth estimation

We tackle mobile manipulation rearrangement tasks from the Behavior-1k dataset, where a robot is required to solve complex tasks requiring exploration, navigation, and manipulation in multi-room environments. The goal is to rearrange objects in a scene based on natural language instructions. We introduce MORE, a novel framework that leverages a large language model to reason about high-level planning, combined with low-level real-world subpolicies.

We study targeted poisoning attacks in a two-agent setting, where an attacker implicitly poisons the effective environment of one agent by modifying the policy of its peer. We provide the computational complexity analysis alongside an approximate algorithm.

We introduce PALMER, a long-horizon planning method that directly operates on high-dimensional sensory input observable by an agent on its own (e.g., images from an onboard camera).