Discovering an informative, or agent-centric, state representation that encodes only the relevant information while discarding the irrelevant is a key challenge towards scaling reinforcement learning algorithms and efficiently applying them to downstream tasks. Prior works studied this problem in high-dimensional Markovian environments, when the current observation may be a complex object but is sufficient to decode the informative state. In this work, we consider the problem of discovering the agent-centric state in the more challenging high-dimensional non-Markovian setting, when the state can be decoded from a sequence of past observations. We establish that generalized inverse models can be adapted for learning agent-centric state representation for this task. Our results include asymptotic theory in the deterministic dynamics setting as well as counter-examples for alternative intuitive algorithms. We complement these findings with a thorough empirical study on the agent-centric state discovery abilities of the different alternatives we put forward. Particularly notable is our analysis of past actions, where we show that these can be a double-edged sword: making the algorithms more successful when used correctly and causing dramatic failure when used incorrectly.

Recently, the FourCastNet Neural Earth System Model (NESM) by (Pathak et al., 2022) has shown impressive results on predicting various atmospheric variables, trained on the ERA5 reanalysis dataset. While FourCastNet enjoys quasi-linear time and memory complexity in sequence length compared to quadratic complexity in vanilla transformers, training FourCastNet on ERA5 from scratch still requires large amount of compute resources, which is expensive or even inaccessible to most researchers. In this work, we will show improved methods that can train FourCastNet using only 1% of the compute required by the baseline, while maintaining model performance or par or even better than the baseline. In this report, we will provide technical details of our methodologies along with experimental results and ablation study of different components of our methods. We have called our improved model FourCastNeXt, in a similar spirit to ConvNeXt (Liu et al., 2022).

Goal-conditioned planning benefits from learned low-dimensional representations of rich, high-dimensional observations. While compact latent representations, typically learned from variational autoencoders or inverse dynamics, enable goal-conditioned planning they ignore state affordances, thus hampering their sample-efficient planning capabilities. In this paper, we learn a representation that associates reachable states together for effective onward planning. We first learn a latent representation with multi-step inverse dynamics (to remove distracting information); and then transform this representation to associate reachable states together in $\ell_2$ space. Our proposals are rigorously tested in various simulation testbeds. Numerical results in reward-based and reward-free settings show significant improvements in sampling efficiency, and yields layered state abstractions that enable computationally efficient hierarchical planning.

Equipping multi-fingered robots with tactile sensing is crucial for achieving the precise, contact-rich, and dexterous manipulation that humans excel at. However, relying solely on tactile sensing fails to provide adequate cues for reasoning about objects' spatial configurations, limiting the ability to correct errors and adapt to changing situations. In this paper, we present Tactile Adaptation from Visual Incentives (TAVI), a new framework that enhances tactile-based dexterity by optimizing dexterous policies using vision-based rewards. First, we use a contrastive-based objective to learn visual representations. Next, we construct a reward function using these visual representations through optimal-transport based matching on one human demonstration. Finally, we use online reinforcement learning on our robot to optimize tactile-based policies that maximize the visual reward. On six challenging tasks, such as peg pick-and-place, unstacking bowls, and flipping slender objects, TAVI achieves a success rate of 73% using our four-fingered Allegro robot hand. The increase in performance is 108% higher than policies using tactile and vision-based rewards and 135% higher than policies without tactile observational input. Robot videos are best viewed on our project website: https://see-to-touch.github.io/.

Teaching dexterity to multi-fingered robots has been a longstanding challenge in robotics. Most prominent work in this area focuses on learning controllers or policies that either operate on visual observations or state estimates derived from vision. However, such methods perform poorly on fine-grained manipulation tasks that require reasoning about contact forces or about objects occluded by the hand itself. In this work, we present T-Dex, a new approach for tactile-based dexterity, that operates in two phases. In the first phase, we collect 2.5 hours of play data, which is used to train self-supervised tactile encoders. This is necessary to bring high-dimensional tactile readings to a lower-dimensional embedding. In the second phase, given a handful of demonstrations for a dexterous task, we learn non-parametric policies that combine the tactile observations with visual ones. Across five challenging dexterous tasks, we show that our tactile-based dexterity models outperform purely vision and torque-based models by an average of 1.7X. Finally, we provide a detailed analysis on factors critical to T-Dex including the importance of play data, architectures, and representation learning.

Online learning via Bayes' theorem allows new data to be continuously integrated into an agent's current beliefs. However, a naive application of Bayesian methods in non stationary environments leads to slow adaptation and results in state estimates that may converge confidently to the wrong parameter value. A common solution when learning in changing environments is to discard/downweight past data; however, this simple mechanism of "forgetting" fails to account for the fact that many real-world environments involve revisiting similar states. We propose a new framework, Bayes with Adaptive Memory (BAM), that takes advantage of past experience by allowing the agent to choose which past observations to remember and which to forget. We demonstrate that BAM generalizes many popular Bayesian update rules for non-stationary environments. Through a variety of experiments, we demonstrate the ability of BAM to continuously adapt in an ever-changing world.