Research
Physical AI, meaning systems that perceive, reason, and act in the real world, is moving robotics from scripted automation toward general-purpose skill. Our work targets what that shift demands when robots operate around people: learning efficiently from human demonstrations and feedback, staying provably safe and compliant during contact, and grounding perception and language in action. We combine machine learning, optimal control, and formal methods, and increasingly build on foundation models and vision-language-action policies through ongoing collaborations. The themes below summarize our current directions and representative results.
Interactive Robot Learning
Robots should acquire skills from the people who already have them. We study how robots learn from demonstrations, corrections, and preferences, and how to keep that learning robust when human data is limited or imperfect. For long-horizon, contact-rich tasks, Impedance Primitive-Augmented Hierarchical RL (ICRA'25) gives a high-level RL policy an action space of parameterized impedance-control primitives, and ST² (RA-M'26) teaches robots long-horizon manipulation by composing skills from sequential human demonstrations. Keypoint parameterization with transportation maps (T-RO'25) generalizes demonstrated motion policies one-shot to novel object configurations. To make imitation robust to the teacher, Beyond the Teacher (ICRA'26) leverages mixed-skill demonstrations to tolerate inconsistent human teachers, RISE (IROSW'25) uses stochastic latent encodings for robust imitation, and PACER (CoRLW'25) curates demonstrations by task progress to curb the compounding errors that destabilize behavior cloning. Beyond demonstrations, DiffusionPack (RA-L'26) uses diffusion models to plan 3D bin packing that respects custom human preferences, and State-Corrected Predictive Preference Learning (ICRAW'26) brings RL and imitation together through preference feedback for robot navigation on uneven terrain. When people correct a robot, Operator-Aware Inversion (CoRLW'26) treats each correction as evidence of what the operator intended, modelling their reaction lag, jitter and bias, so that frozen generative policies adapt from more accurate targets. We also co-organized the 3rd Workshop on Human-Interactive Robot Learning (HRI'24).
Safe and Compliant Control
Robots acting in the physical world, and especially in contact with people, must be compliant and provably safe, with guarantees that hold during interaction and not just on average. We bring formal methods and control-theoretic certificates together with learning to enforce this. Much of our work learns safety directly from offline data, without risky online exploration: Bridging Safety and Performance (NeurIPS'26) uses offline reinforcement learning to reconcile safety constraints with task performance in autonomous systems, Safe Flow Q-Learning (RLC'26) couples reachability analysis with flow-based policies for offline safe RL, and V-OCBF (TMLR'26) learns value-guided control barrier functions that act as safety filters. For physical human-robot interaction, SafeDMPs (ICRA'26) embeds Signal Temporal Logic (STL) safety specifications directly into Dynamic Movement Primitives, so that adaptive motions remain provably safe during physical contact, while certified RL for variable impedance control (ICRA'26) introduces Certified Gaussian-Manifold Sampling (C-GMS), which keeps every impedance profile sampled during RL within Lyapunov-based stability certificates, so impedance is modulated optimally with guaranteed stability. Without C-GMS, unconstrained exploration drives the robot past its limits and triggers a hard stop, as the video shows.
Foundation Models for Robotics
Foundation models, including large vision-language models (VLMs) and, increasingly, vision-language-action (VLA) models pretrained on web- and robot-scale data, supply semantic priors and broad generalization that can ground robot behavior in natural language and perception. We study how to turn these priors into natural, flexible human-robot interaction. Our work on OVITA (RA-L'25) harnesses pretrained foundation models to adapt reference trajectories from open-vocabulary language instructions at execution time, producing interpretable, parameterized edits to robot motion without any task-specific retraining. Building on VLA models directly, Gripper-aware Vision Language Action Models (ECCV'26) make the policy aware of the robot's end-effector, improving grasping and manipulation across diverse gripper embodiments. We also examine how these models are best composed into robot systems: Where Reasoning Does Not Matter (RSSW'26) isolates the contribution of chain-of-thought reasoning in a modular, agentic VLM manipulation pipeline, and Latent Dynamics Modulation Guidance (ICRAW'26) steers diffusion policies away from semantically specified obstacles.
3D Vision, SLAM and World Models
Acting in unstructured environments demands spatial understanding from 3D reconstruction, simultaneous localization and mapping (SLAM), and world models that capture both geometry and semantics. We are building active-perception pipelines that tie real-time scene reconstruction and object-level mapping to manipulation planning. A first step in this direction, AffordMatcher (CVPR'26), learns object affordances in 3D scenes from visual signifiers, connecting 3D scene understanding to where and how a robot should act in complex, cluttered spaces. Complementing this, Beyond Visual Grasping (IROS'26) benchmarks the full pipeline of complex grasping from detection to execution, exposing where perception-to-action systems break down on challenging, real-world grasps. Toward world models of how objects move, GenieMotion (CVPRW'26) learns object-centric 4D motion representations from a single image using generative demonstrations.
Optimization and Optimal Control
Optimization and optimal control offer principled, data-efficient ways to synthesize controllers and motion plans, often with stability or optimality guarantees. Our Adaptive Critic (T-CST'24) framework uses approximate dynamic programming with neural value-function approximation to learn near-optimal controllers for uncertain manipulators online, without explicit system identification. For motion planning, The Topological Cage (ICRAW'26) uses signed distance fields and the medial axis of free space to find near-optimal paths, of the kind roadmap planners such as PRM produce, without their sampling-based search.
