Cho Hyeon Joon

Research

Four projects on robot learning for physical manipulation, 2024–2026. Each page states the problem, my role versus the team's, the method, the evaluation condition behind every number, and the limitations.

Recovery-Aware LPB for RMP-Induced OOD States

2026 · Manuscript in preparation

Reactive safety layers such as RMPflow keep a manipulator away from obstacles, but the avoidance motion can leave the robot in states that never appeared in the demonstrations. I designed the recovery bridge and evaluated the proposed method in 160 RB10 trials. It completed 146 trials (91.3%); the original LPB reached 64.4% under the same intervention setting.

91.3% Recovery-aware LPB. 146 successes in 160 trials of the proposed method across eight OOD start configurations.

Force-Feedback Teleoperation for Real-Robot Imitation Learning

RAGTAL / DNA-HERO Team Project

2025 · Completed

The DNA-HERO team built a custom teleoperation system with external force/torque sensing and force feedback for collecting manipulation demonstrations on an RB10. As Learning Lead, I built the demonstration-to-rollout workflow, curated the training data, and conducted 50 real-robot rollouts for each evaluated model.

9/10 Integrated final trials. Successful demonstrations during the final presentation setting.

Grounded Language-to-Motion Planning with Memory and Gripper Constraints

Extension of LMTG

2025 · Evaluated on 15 scenarios

Language-model planners for tabletop manipulation often fail for reasons unrelated to reasoning: the object named in the instruction is not grounded in the image, the chosen grasp does not fit the gripper, or the same mistake is repeated after a failed attempt. I connected visual grounding and execution memory to the planner, then evaluated the system on a fixed set of 15 scenarios. Success increased from 9/15 to 12/15.

12/15 Improved system. Successes in the defined 15-scenario evaluation.

Zero-Shot Language-to-Motion Planning with Geometric Validation

Reproduction and extension of LMTG

2024–2025 · Prototype · early robot trials

LMTG generates manipulation trajectories from natural-language instructions without task-specific training, but a language model has no geometric model of the workspace, so generated waypoints can pass through objects or approach with an unusable end-effector orientation. I implemented the task representation, waypoint parser, and collision checks, then connected the planner to simulation and initial real-robot trials.

Zero-shot Planning setting. No task-specific model fine-tuning in the prototype.