Research

Research Interests

Scalable Robot Manipulation

• Data efficiency and generalization in manipulation
• Learning across heterogeneous data distributions
• Sim-Real-Generated data integration

Representation-Centric Learning

• Action-defining representations instead of policy learning
• Progress-based, articulation-aware representations
• Scale-invariant and structure-preserving embeddings

Articulated Object Manipulation

• Drawer, door, cabinet, articulated tools
• Metric goals (e.g., “open 5cm”, “30% progress”)
• Joint state, articulation constraints, motion limits

Simulation as a Data Engine

• Simulation for supervision, not realism
• Joint values, articulation progress, motion constraints
• Sim-and-Real Co-Training paradigms

Video & Synthetic Data

• Video generation for observation diversity
• In-the-wild visual distribution expansion
• Limits of generated video for action supervision

3D Object Flow & Part-level Motion

• Part-level motion over absolute pose
• Allowed motion subspace modeling
• Flow-based articulation reasoning

Research Themes

Theme 1: Sim-Real-Generated Data as a Unified Distribution

Simulation data improves real-world performance even when visually mismatched. The key factor is structured, action-relevant supervision. The focus is on distribution expansion, not gap minimization.

Theme 2: Quantitative Articulation Representation

Current models cannot execute metric language goals. I propose articulation progress as a learnable target. Simulation enables otherwise impossible annotations.

Theme 3: Video as Observation Augmentation

Generated video cannot provide reliable action labels. It is used to expand visual and environmental diversity. Action is inferred via structured constraints, not direct regression.

Theme 4: Flow-based Motion Understanding

Articulated objects are defined by motion constraints. Flow provides relative, constraint-aware motion cues. It is used as auxiliary signal, not primary supervision.

Research Statement

My research focuses on scalable robot manipulation by shifting the learning target from end-to-end policies to structured representations that define action. I study how simulation, real-world data, and generated observations can be integrated into a unified learning distribution through articulation-aware, metric, and scale-invariant representations. By treating simulation as a source of structured supervision rather than visual realism, and video generation as a tool for observation diversity rather than action labeling, my work aims to enable robust manipulation of articulated objects under diverse and previously unseen conditions.

Selected Topics

• Sim-and-Real Co-Training
• Diffusion Policies (data-centric view)
• Articulated Progress Estimation
• Representation Design for Action
• Synthetic Data Pipelines