Deployed diffusion policies utilizing FiLM-ResNet and temporal U-Net for real-time 1 kHz control of a robotic arm, achieving 92.14% test-time accuracy.
Collected and structured over 130 RGB-D frames and action trajectories into ReplayBuffers, enabling robust training and evaluation of robotic control models.
Improved action diversity by 35% over IBC and BET baselines using multi-modal sampling with Diffusion Policy, enhancing robotic arm adaptability.
Modified the evaluation script from an RTDE-based interface to the ROS2 Framework, resulting in 92% test-accuracy for the Franka arm and streamlining research workflows.