Modelpublic

siyulw2025/aic-rlt-cable-insertion

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

Model Card

AIC RLT — SFP Cable Insertion

RLT (RL Token) checkpoint for the AIC cable insertion challenge.

What's inside

File	Description
`phase1_rl_token.pt`	Phase 1: RL token encoder pretrained on XVLA (lerobot/xvla-base) embeddings extracted from 100 synthetic demo episodes. Loss: 0.65 → 0.024 over 50 epochs.
`phase2_offline.pt`	Phase 2: Actor + Critic trained offline via TD3+BC on 13,100 demo transitions with synthetic distance-to-goal rewards. 5,100 gradient steps.

Architecture

—RL token encoder: Perceiver-style cross-attention, compresses (115, 1024) XVLA embeddings → 2048-dim z_rl
—Actor: 3-layer MLP (256→256→action), outputs action chunks (C=10, 7-dim TCP)
—Critic: Twin Q-networks for TD3-style pessimistic value estimation

Training data

—100 synthetic episodes of scripted cable insertion (generate_synthetic_data.py)
—XVLA embeddings extracted from Florence-2 encoder (lerobot/xvla-base)

Usage (ROS 2 simulation)

bash

pixi run ros2 run aic_model aic_model \
  --ros-args \
  -p use_sim_time:=true \
  -p policy:=aic_example_policies.ros.RunRLT \
  -p policy_args.checkpoint_path:=$(hf_hub_download siyulw2025/aic-rlt-cable-insertion phase2_offline.pt)

See aic_example_policies/ros/RunRLT.py for the full inference policy.