Atharva Bhorpe · Project
Autonomous Insertion Challenge
Team-based imitation learning for fiber-optic connector insertion with a simulated UR5e arm.
- Role
- Policy tuning and data collection in a team project
- Tools
- LeRobot · PyTorch · Diffusion Policy · Gazebo
Problem and objective
Train imitation-learning policies to insert SFP and SC fiber-optic plugs with a simulated UR5e arm in Gazebo for Intrinsic’s AI for Industry Challenge.
My contribution and approach
The supplied résumé describes tuning a Diffusion Policy with per-camera ResNet34 encoders, a wider U-Net and an extended action horizon.
It also records DAgger data collection to improve robustness to distribution shift and model training on the RWTH HPC cluster. This was a team project, not an individual challenge entry.
Reported results and limitations
The résumé reports a score of 86/100 on the SFP subtask, a result above an ACT baseline, and a 35th-place team finish in the global challenge.
These are author-reported results. No public leaderboard record, evaluation logs or code link was supplied for independent verification. The work concerns simulated insertion; it does not establish physical-robot performance.
Sources and evidence
Source: supplied résumé. Public code, challenge evaluation and demonstration links have not been supplied.