aleleanza/diffusion-policy-vac-fridge

Diffusion Policy — VAC Fridge (input/output ablation)

Eight robomimic Diffusion Policy checkpoints for the variable-impedance pick ("fridge") task on a Doosan M0609 + Inspire RH56 hand under a Variable Admittance Controller (VAC). Trained on aleleanza/vac-fridge-single-cam (200 episodes, 54,833 frames @ 20 Hz, single RGB camera).

This is the fridge-task counterpart of aleleanza/diffusion-policy-vac-pipe: an ablation over two output representations (relative to the admittance filter) × input modality, plus a larger-UNet variant per side.

Variants (subfolders)

Subfolder	Output (action)	Inputs (obs)	Action dim	Best val loss	@epoch
vac_preimg	pre: user_cmd[6] + K + ζ + hand_binary	vision	9	0.11807	240
vac_preimg_state	pre	vision + state	9	0.09940	240
vac_preimg_state_wrench	pre	vision + state + wrench	9	0.10013	597
vac_preimg_state_wrench_big	pre (larger UNet)	vision + state + wrench	9	0.10529	275
vac_postimg	post: vel_cmd[6] + hand_binary	vision	7	0.10254	240
vac_postimg_state	post	vision + state	7	0.09851	193
vac_postimg_state_wrench	post	vision + state + wrench	7	0.09467	193
vac_postimg_state_wrench_big	post (larger UNet)	vision + state + wrench	7	0.10136	263

• pre = predict the operator command before admittance, including the compliance command itself (stiffness_cmd K + damping ζ) → the policy learns to set compliance.
• post = predict the Cartesian velocity executed after admittance → bypasses the admittance filter and drives velocity directly.

Each subfolder contains: best.pth, last.pth, config.json, action_stats.json (action min-max normalization + components + hand binarization), dataset_summary.json (train/valid split).

Architecture & training

• Algorithm: robomimic Diffusion Policy (DDPM noise-prediction loss).
• Backbone: conditional UNet [128, 256, 512]; the _big variants use [256, 512, 1024] (~91.8M params vs ~89.4M).
• Horizons: observation 2, action 4, prediction 8 (frame stack 2, seq length 8).
• Image: single camera → front_rgb, 84×84.
• Diffusion: DDPM 50 train/infer steps (DDIM 10-step configurable for faster inference).
• Schedule: up to 600 epochs, batch 16, lr 1e-4.
• state is built at runtime from current TCP + Inspire hand joints; wrench from /bota_ft_sensor/wrench. Hand head is binarized (hand_open_binary).

Reading the results

Validation losses are tightly clustered (~0.095–0.118). Adding state helps both pre and post; the best overall is vac_postimg_state_wrench (0.0947). The larger-UNet (_big) variants did not improve validation at this dataset size. Post (vel_cmd) targets are marginally easier than pre (user_cmd + K + ζ), consistent with the pipe ablation though the gap here is much smaller.

Inference (ROS 2)

Run with the project's robot_learning real-time inference nodes (Doosan M0609 + Inspire hand). The action contract determines the runner:

• pre* variants (9D) → diffusion_policy_vac_preimg_runner: publishes action[:6]→/delta_pose_cmd, [6]→/predicted_K, [7]→/predicted_zeta, [8]→/inspire_hand/left/cmd; consumed by variable_admittance_node (variable_K:=true).
• post* variants (7D) → diffusion_policy_fixed_k_runner in vel_cmd mode: streams velocity directly via the DSR speedl interface (no admittance node).

Pre-wired launchers exist under robot_learning/scripts/ (e.g. launch_vac_inference.sh preimg_state_wrench); *_wrench* variants additionally need the Bota driver. action_stats.json provides the exact normalization to undo at inference.

Intended use & limitations

• Use: research on force-aware / compliance-predicting imitation learning; VAC baselines.
• Limitations: single task, single embodiment (M0609 + RH56), single 84×84 camera, binary hand. Validation differences between variants are small. Not validated for safety-critical or autonomous deployment.

Related

• Dataset: aleleanza/vac-fridge-single-cam
• Pipe counterpart: aleleanza/diffusion-policy-vac-pipe
• Collection: VAC — Fridge