Research

My research interests lie in robot learning systems that combine data-driven policies with structure from robotics, control, and physical interaction.

Main Research Themes

Long-Horizon Imitation Learning

I am interested in how robot policies can learn temporally extended manipulation skills from demonstrations. This includes behavior cloning, Action Chunking Transformers, VQ-style action abstractions, diffusion policies, and recovery-aware data collection.

Reinforcement Learning under Partial Observability

Many robotic tasks are partially observable because the robot only has local sensors, noisy perception, or incomplete state information. I study recurrent policies such as PPO-LSTM and PPO-GRU for such settings.

Vision-Language-Action Models

I am interested in how high-level language-conditioned reasoning can connect to low-level continuous robot control through action decoders, proprioception, temporal modeling, and grounding in real demonstrations.

Practical Robot Data Collection

A major bottleneck in robot learning is collecting useful, trustworthy, and scalable data. I am exploring low-cost data collection pipelines using SO101 arms, cameras, teleoperation, and structured evaluation.

Current Questions

• How can ACT-style policies be made more robust to disturbances?
• When should we combine imitation learning with motion planning or RMPFlow-style smoothing?
• How much data is needed for reliable long-horizon manipulation on low-cost hardware?
• How can recurrent RL policies form useful memory in sparse-reward POMDPs?
• How should VLA models represent actions for precise robot control?