disentangled-vla/libero-island-ablation-v9-infonce-low

V9 — V7 with InfoNCE weight reduced 125× so L1 action dominates 2:1

Ablation V9 of the Stage3 Qwen2.5-VL framework: identical to V7 (single AttentiveLatentHead A + L1 + InfoNCE with our stage1 task target) BUT latent_infonce_weight reduced from 1.0 (V7) to 0.008.

In V7 the InfoNCE contribution (≈3.64) dominated the action L1 (≈0.056) at 65:1; cos(z_a, z_target) stalled at ~0.07 and InfoNCE never effectively pulled z_a toward the target. V9 inverts the balance so action contribution (0.056) is 2× the InfoNCE contribution (0.029), giving the action signal the lead while keeping InfoNCE as a gentle alignment pressure.

Evaluated on the LIBERO-Spatial island viewpoint protocol from Xing et al. 2025 (CoRL) "Shortcut Learning in Generalist Robot Policies".

Architecture (identical to V7)

• Backbone: Qwen/Qwen2.5-VL-3B-Instruct + LoRA (r=32, q/v_proj)
• Single AttentiveLatentHead A (8 queries, depth 2, 16 heads, proj_dim=4096)
• ResNetActionHead (hidden 2048, 2 blocks, chunk=16, action_dim=32 padded to 7-D)
• V-JEPA target: our stage1 ViT-L + LoRA, 4-quad attentive future (same as Ours / V7)
• Loss: 1.0·L1_action + 0.008·InfoNCE(z_a, z_target_ours)
• No SIGReg / no Head B / no domain branch / no augmentation

Training

• Data: LIBERO-Spatial island A+B paper-200 protocol (~211K frames, 1,724 eps)
• bs=64 single GPU, lr=1e-4 cosine, 100K steps, 10 epochs

Evaluation protocol

6 settings × 5 tasks × 5 trials = 25 episodes / setting.

Setting	task subset	viewpoint	type
AA	{0,1,3,5,8}	[0.20, 0.35]	in-domain A
BB	{2,4,6,7,9}	[0.65, 0.80]	in-domain B
C_A	{0,1,3,5,8}	0.50 fixed	center, A-tasks
C_B	{2,4,6,7,9}	0.50 fixed	center, B-tasks
AL_B	{2,4,6,7,9}	0.20 fixed	OOD (B-tasks @ A-leftmost)
BR_A	{0,1,3,5,8}	0.80 fixed	OOD (A-tasks @ B-rightmost)

Results — epoch_10

Setting	SR
AA	0.68
BB	0.72
C_A	0.08
C_B	0.28
AL_B	0.0
BR_A	0.08

Files

eval_ablation_v9_first_videos/
├── {setting}_results.jsonl
├── {setting}_summary.json
└── videos/{setting}/task{N}_trial{T}_a{angle}_{OK,FAIL}.mp4

Videos: 25 ep × 6 settings = 150 mp4 total.