580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can robots learn effective motor skills through self-supervised and visual observation-driven imitation?
This research theme focuses on enabling robots and agents to acquire goal-directed motor skills by observing demonstrations visually, often without explicit action labels. It addresses challenges related to learning from sparse or unlabeled demonstrations, self-supervised skill discovery, and goal-conditioned policy learning. This is essential for scalable robot learning, where human demonstrations are easier to provide as visual observations rather than hand-labeled action sequences. The approaches investigated here emphasize learning internal representations and policies that connect current observations to goals by distilling skills from the robot’s autonomous environment exploration.
Key finding: This paper introduces a framework where a robot learns goal-conditioned skill policies (GSPs) through self-supervised exploration without expert supervision. The GSP predicts action sequences that take the agent from current... Read more
Key finding: The authors propose an imitation learning system training a deep neural network controller from a Belief Space Planning (BSP) algorithm’s output instead of a human expert. This approach enables a low-cost, low-accuracy... Read more
Key finding: This work addresses the challenge of self-initiated imitation, where a robot autonomously determines which behaviors within a continuous sensory stream are worth imitating without explicitly segmented demonstrations. The... Read more
Key finding: The paper presents a model enabling robots to predict and imitate others’ motions by reusing their own forward-inverse body model, incorporating viewpoint translation and triadic relations (infant, parent, object). Using... Read more
2. How can imitation models incorporate hierarchical segmentation and recognition to enable modular and scalable learning of complex motor sequences?
This theme investigates methods for decomposing complex observed behaviors into libraries of reusable movement primitives, facilitating segmentation, recognition, and reproduction of constituent action units. This hierarchical approach supports scalability in imitation by building on pre-learned primitives to recognize and generate complex actions. Such a modular framework addresses challenges related to temporal smoothing of primitives, detection of unknown primitives, and the simultaneous requirement of segmentation for recognition versus reproduction. It holds significance for robots imitating human tasks requiring multi-step execution.
Key finding: The paper presents a framework for segmenting and recognizing observed complex human movements by leveraging a pre-existing library of movement primitives. The approach detects temporal boundaries in smoothed sequences and... Read more
Key finding: This work develops a formal mathematical framework to optimize imitation by automatically learning which features (endpoints, trajectories, joint angles) to imitate based on their invariance. The hierarchical approach... Read more
Key finding: The authors explore the issue of ‘what’ to imitate—whether exact actions, effects, or goals—in particular for heterogeneous embodiments between demonstrator and imitator. The study emphasizes the importance of task... Read more
3. What are effective frameworks and evaluation methodologies for imitation learning systems to develop human-like and context-aware agent behaviors?
This research theme encompasses methodological advances in designing deterministic and hybrid imitation/reinforcement learning systems, as well as approaches to objectively assessing the human-likeness and plausibility of agent behaviors. It includes rule-based and physics-guided learning models, integration of cognitive architectures, and metrics capturing context-informed behavioral similarity. Addressing evaluation and design frameworks is critical for human-robot interaction, autonomous driving, and social acceptance of imitation-based agents.
Key finding: This paper proposes a hybrid training scheme that combines reinforcement learning (RL) and imitation learning (IL) to develop agents that surpass pure IL in task performance while exhibiting more human-like behavior than... Read more
Key finding: This research develops a multidimensional, context-aware metric to objectively assess the plausibility and human-likeness of artificially generated driving trajectories. By categorizing situations contextually and comparing... Read more
Key finding: This work introduces deterministic AI design frameworks leveraging hierarchical imitation models as alternatives to neural networks. Emphasizing accuracy, correctness, and personalization without relying on big data, it... Read more
Key finding: This study integrates a psychological cognitive architecture (ACT-R) with a virtual human avatar to simulate human perception, cognition, and action coherently during task execution. Specifically, extensions to perception and... Read more