aic-port-insertion-policies

AIC Port Insertion — final checkpoints (bachelor's thesis, ETF 2026)

Final checkpoints from all 6 imitation-learning runs for the task of a UR5e robot inserting an SFP cable into a port (Isaac Lab simulation, LeRobot trainers with an auxiliary phase head). Author: Mihajlo Stevanović. Code, tutorials, and env: https://github.com/etf-robotics/etf_robotics_aic

Each checkpoint contains pretrained_model/ (config + safetensors — enough for eval) and training_state/ (optimizer state — for --resume).

Results (N=150, eval seeds 0–2, 3 mm success threshold)

• DP full: 1.3 ± 1.9 % success; ACT full: 0.7 ± 0.9 %.
• Median closest approach (tip→seat) plateaus at ~47 mm across all runs — the bottleneck is the final contact phase (force-free DiffIK), not the model or the data.
• Recording of a successful DP episode: https://youtu.be/SS5MC-dcSX4

Evaluation

docker exec -w /workspace/isaaclab isaac-lab-base \
  ./isaaclab.sh -p etf_robotics_aic/scripts/eval_demos.py \
  --headless --enable_cameras \
  --ckpt etf_robotics_aic/outputs/dp_phase_full_001/checkpoints/200000/pretrained_model \
  --n_episodes 50 --seed 0 --save_videos

Checkpoints are restored to the expected paths with snapshot_download(..., local_dir='etf_robotics_aic/outputs') — full instructions: docs/06_artifacts_and_archive.md in the GitHub repo.

Companion eval artifacts (metrics + recordings): https://huggingface.co/datasets/Mihajlo04/aic-port-insertion-eval Training curves: wandb entity sm220315d-etf-, projects aic-{act,dp}-phase-port-insertion.