Towards Intuitive Human-Robot Collaboration: From Human Monitoring to Robot Learning

Artificial intelligence has made remarkable progress in digital domains, yet robots still struggle to achieve the physical intelligence required for seamless operation in everyday environments. This thesis advances the vision of intuitive human-robot collaboration by addressing two complementary challenges: real-time coordination with human partners in structured settings, and robust learning for general-purpose robotic skills in unstructured ones. The first part focuses on enabling robots to anticipate and synchronize with human actions at the level of individual tasks. We introduce a novel online Dynamic Time Warping variants for real-time progress estimation, coupled with methods for predicting action completion times. Building on these techniques, we develop the Proactive Assistance through action-Completion Estimation (PACE) framework, which aligns robotic support with human workflows in collaborative assembly. Experiments with human participants demonstrate that this approach reduces idle times, improves fluency, and fosters trust, highlighting the central role of proactive timing in making robots more natural partners. The second part explores how vision-language-action (VLA) models can extend robotic capabilities beyond scripted tasks to flexible, instruction-driven behaviors. We propose Point0, a VLA that augments the pre-trained π0 backbone with point-cloud features from a single external RGB-D sensor. This design eliminates the reliance on wrist-mounted cameras while enhancing spatial reasoning, yielding significant gains on the LIBERO benchmarks. Our results confirm that explicit 3D geometry improves robustness to occlusion and generalization across manipulation tasks, while also revealing current limitations: the scarcity of large-scale depth datasets, the biases of 2D-centric pretraining, and the need for real-world validation. Together, these contributions move toward a future in which robots act as intuitive collaborators—anticipating human needs in structured settings while flexibly adapting to novel tasks in open-ended environments. By bridging advances in real-time human monitoring with foundation-scale learning, this thesis takes a step closer to robots that can complement human strengths, adapt to human actions, and operate as trusted partners in homes, workplaces, and beyond.