On Generalization and Benchmarking on Physical Robots

Abstract

Robot learning research has seen significant advancements. However, the field remains predominantly demo-driven, making direct comparisons between methods difficult without replicating them on individual setups. This places a substantial burden on making scientific progress. While many simulation benchmarks exist, they usually feature contrived datasets and do not accurately reflect real-world performance. The absence of widely recognized benchmarks and real-world applicability makes it difficult to ascertain scientific advancements.

In my thesis, I propose two works that tackle these challenges. In Chapter 2, instead of assuming access to datasets of any quality, we suggest that near-optimal and safe demonstrations collected from out-of-domain tasks are more practical data-sources for real world robot learning. We further propose a set of experiments that evaluate the generalization capabilities of offline reinforcement learning (ORL) and imitation learning methods within this framework. Our study finds that ORL and imitation learning prefer different action spaces, and that ORL algorithms can generalize from leveraging offline heterogeneous data sources and outperform imitation learning.

The second work, introduced in Chapter 3, presents an initiative towards establishing a real-robot benchmark: shared tasks and robots for evaluation that are remotely submitted to and an open-source dataset in this setting. Our benchmark suite includes common manipulation tasks that require challenging generalization to unseen objects, positions, and lighting. Initial results from the benchmark and the launch of a NeurIPS competition highlight the feasibility of such systems.