Autumn Kwon

This project aims to develop a safe, multi-step robotic manipulation framework using an ORCA dexterous hand mounted on a robot arm. The proposed approach combines imitation learning for pick-and-place operations with reinforcement learning for contact-rich tool manipulation, using furniture assembly as a case study. A shared safety controller is designed to enforce motion constraints, collision checks, and visual progress verification. The project investigates how complementary learning methods and safety mechanisms can enable reliable manipulation in both simulated and real-world environments.

This project investigates how user authority influences sycophantic behavior in LLMs during multi-turn reasoning dialogues. We compare instruction-tuned and reasoning-oriented Qwen3 models under controlled dialogue settings, varying both the strength and temporal ordering of user authority. Using trajectory-based metrics, we analyze when and how frequently models change their responses under repeated user pressure. Our findings show that reasoning-oriented models generally demonstrate greater resistance to sycophancy, but remain vulnerable to strong authority cues. In particular, introducing strong authority early in a conversation leads to earlier and more frequent answer reversals, highlighting the importance of interaction dynamics in LLM reliability.

This project investigates symbolic feedback (SF) as a mechanism for incorporating interpretable behavioral structure into reinforcement learning for physics-based motion imitation. We introduce phase-based symbolic signals derived from motion alignment, stability, and temporal consistency, augmenting conventional imitation rewards without modifying the underlying policy architecture. Using the MimicKit framework, we evaluate symbolic feedback with PPO and AWR across challenging humanoid motion tasks under stochastic perturbations. Our experiments show that while improvements in cumulative rewards vary across tasks and algorithms, symbolic feedback consistently encourages progression toward higher-level behavioral phases, producing more structured and interpretable motion patterns.

This research investigates visual domain randomization for improving the sim-to-real transfer of reinforcement learning policies in autonomous racing. Eight autoencoder-based visual representations were trained under different lighting conditions, including variations in light direction, intensity, and shadows. The models were evaluated in the DonkeyCar simulator and transferred to a physical JetRacer on a figure-eight track. Experimental results showed that combining light-direction and shadow randomization improved real-world driving robustness compared with the baseline, particularly under changing illumination conditions. Additional experiments explored combining visual randomization with sensor and action randomization.