Tabitha Edith Lee

Due to fundamental exploration challenges without informed priors or specialized algorithms, agents may be unable to consistently receive informative rewards, leading to inefficient or intractable learning. To address these challenges, we introduce CURATE, an automatic curriculum learning algorithm for reinforcement learning agents in structured task spaces of monotonic difficulty. Through "exploration by exploitation," CURATE dynamically scales the task difficulty to match the agent’s current competence. By exploiting its current capabilities that were learned in easier tasks, the agent improves its exploration in more difficult tasks. Our key insight is that the learning improvement in tasks that are close to those used for training is inversely proportional to their difficulty, and an agent that chooses a nearby distribution of the easiest unsolved tasks at any given time can automatically induce an easiest-to-hardest curriculum in these task spaces. To achieve this, CURATE conducts policy search in the task space to learn the best task distribution for training. As the agent’s mastery grows, the learned curriculum adapts in an approximately easiest-to-hardest and task-directed fashion, efficiently culminating in a performant agent. Our experiments across three diverse domains (MiniGrid, Procgen, BipedalWalker) demonstrate that CURATE learns effective curricula for sample efficiency and proficiency with the potential for yielding broadly capable agents, matching or exceeding prior curriculum methods that do not require informed initialization or predefined schedules.

Two decades into the third age of artificial intelligence, the rise of deep learning has yielded two seemingly disparate realities. In one, massive accomplishments have been achieved in deep reinforcement learning, protein folding, and large language models. Yet, in the other, the promises of deep learning to empower robots that operate robustly in real-world environments have yet remained unfulfilled. Vast diversity of objects, distribution shifts, long-tailed phenomena: outside of laboratories, real-world environments challenge modern statistical learning assumptions of the data.

Although such environments have generally been referred to as "unstructured," this terminology belies their nature. Real-world environments are not "unstructured," but arise because of structure: the underlying causal processes that generate the observed data. In this view, robots should not only reason and learn with respect to data, but also the data generating processes. Such processes can be formalized by the language of causality. Therefore, to learn and leverage the structure of these "open-world" environments, new causal-based robot learning algorithms are needed.

Towards this goal, this thesis explores a diversity of robot learning problems, from perception to control. First, we explore how perception models can be learned using sim-to-real transfer from synthetic data (DREAM, FormNet). In these works, relevant features were learned through domain randomization, leading to insights into how structure can be learned more directly using causality. To this end, we introduce structural sim-to-real transfer, where simulation can serve as a causal reasoning engine for the robot to select the relevant features for a control policy (CREST) or skill (SCALE). By construction, these policies are robust to irrelevant distribution shifts that would otherwise stymie correlation-based deep learning. Next, the rich interplay between control, dynamical systems, and causality is explored through the Learning By Doing (LBD) competition and LMeshNet, a methodology for constructing hybrid causal world models that integrate both latent and semantic information. Lastly, we examine curriculum learning. In ACL, we explore the commonalities and differences between human and agent curriculum learning. Then, we employ these lessons learned for CURATE: how agents can manipulate the sequencing of training data to efficiently learn a control policy to solve a desired, difficult task.

The advantages of adopting the principles of causal inference have been witnessed to date in fields such as biomedical sciences, economics, and genomics. In the machine learning community, it has recently been argued that such principles should be integrated to harness deep learning, towards causal learning of representations. Analogously, this thesis forwards that the robot learning community stands to gain by leveraging the principles of causality. In so doing, this new paradigm holds promise for robots to learn and leverage structure within the open world through causal robot learning for manipulation.

We propose SCALE, an approach for discovering and learning a diverse set of interpretable robot skills from a limited dataset. Rather than learning a single skill which may fail to capture all the modes in the data, we first identify the different modes via causal reasoning and learn a separate skill for each of them. Our main insight is to associate each mode with a unique set of causally relevant context variables that are discovered by performing causal interventions in simulation. This enables data partitioning based on the causal processes that generated the data, and then compressed skills that ignore the irrelevant variables can be trained. We model each robot skill as a Regional Compressed Option, which extends the options framework by associating a causal process and its relevant variables with the option. Modeled as the skill Data Generating Region, each causal process is local in nature and hence valid over only a subset of the context space. We demonstrate our approach for two representative manipulation tasks: block stacking and peg-in-hole insertion under uncertainty. Our experiments show that our approach yields diverse skills that are compact, robust to domain shifts, and suitable for sim-to-real transfer.

We present CREST, an approach for causal reasoning in simulation to learn the relevant state space for a robot manipulation policy. Our approach conducts interventions using internal models, which are simulations with approximate dynamics and simplified assumptions. These interventions elicit the structure between the state and action spaces, enabling construction of neural network policies with only relevant states as input. These policies are pretrained using the internal model with domain randomization over the relevant states. The policy network weights are then transferred to the target domain (e.g., the real world) for fine tuning. We perform extensive policy transfer experiments in simulation for two representative manipulation tasks: block stacking and crate opening. Our policies are shown to be more robust to domain shifts, more sample efficient to learn, and scale to more complex settings with larger state spaces. We also show improved zero-shot sim- to-real transfer of our policies for the block stacking task.