Hyeonjun Cho

Research · 2025 · Completed

Force-Feedback Teleoperation for Real-Robot Imitation Learning

RAGTAL / DNA-HERO Team Project

An RB10 imitation-learning system trained from force-feedback teleoperation demonstrations and evaluated through repeated real-robot rollouts.

Problem

The DNA-HERO team built a custom teleoperation system with external force/torque sensing and force feedback for collecting manipulation demonstrations on an RB10. The learning question was whether demonstrations gathered through that system could train a policy that completes the team's three manipulation tasks on the real robot, and which policy class should be deployed.

Key result

9/10

Integrated final trials. Successful demonstrations during the final presentation setting.

Contribution

As Learning Lead, I built the demonstration-to-rollout workflow, curated the training data, and conducted 50 real-robot rollouts for each evaluated model.

My Role

  • Led learning-system integration from teleoperation demonstrations to policy rollout.
  • Collected and curated real-robot demonstrations and ran 50 rollouts per evaluated model.
  • Diagnosed failure cases and coordinated model iteration with the hardware team.

Team Contribution

  • Team members designed and assembled the force-feedback teleoperation hardware and supporting electronics.
  • Hardware, control, learning, and presentation work were integrated across the full DNA-HERO team.

Method

As Learning Lead I owned the pipeline from demonstration preprocessing to rollout: curating the teleoperation data, training BC-Transformer and image-based diffusion policies, deploying checkpoints on the robot stack, and running RB10 rollouts. The deployed stack integrated dual RGB-D observations, OnRobot gripper I/O, and RB10 control on top of the team's teleoperation hardware.

Diagram showing force-feedback teleoperation demonstrations flowing into dataset curation, policy learning, and RB10 robot rollout evaluation.
Figure 1. I led the learning loop. The team built and integrated the teleoperation hardware, controls, and robot platform.
  1. Force-feedback teleoperation
  2. Demonstration collection
  3. Imitation-learning policy training
  4. RB10 rollout evaluation

Evaluation results

ConditionResultNote
Integrated final trials9/10Successful demonstrations during the final presentation setting.
Rollouts per model50Real-robot evaluation runs used for model comparison.
DNA Capstone DesignGrand PrizeAwarded for the integrated team project.

BC-Transformer and diffusion policy were compared in 50 real-robot rollouts per model. The integrated system completed 9 of 10 trials across three manipulation tasks in the final-presentation setting; this result describes that presentation condition rather than a broad benchmark. The project received the Grand Prize at the Sungkyunkwan University stage of the DNA-HERO industry-academia program.

Experiment evidence

Video 1. Leader-arm teleoperation used to collect demonstrations for three task configurations.
Video 2. Leader–follower gripper synchronization check before demonstration collection. This verifies command coupling; it is not a learned-policy rollout.
Video 3. Representative real-robot rollouts from the three final-presentation task families. The reported 9/10 result is limited to that integrated presentation setting.

Limitations

  • The 9/10 result reflects the final-presentation condition, not a broad benchmark.
  • The hardware team designed and fabricated the teleoperation system.

Materials

Stack. Python, PyTorch, ROS 2, RB10, Teleoperation, Imitation Learning.

Code and materials. Request team repository access