easyminnn/atq-labelgated-2.5x-vlm_contact-disc611-60k

ATQ label-gated MoE — compressed speed 2.5x, label vlm_contact, discrete dims fixed, 60k steps

Inference-only checkpoint (weights + config; optimizer/scheduler/RNG state stripped). Fine-tuned from nvidia/GR00T-N1.5-3B on RoboCasa MG 300 (7,200 episodes, 3 cameras 256x256, 20 fps) with VLM speed-ratio labels baked into the action tensor (ratio_label, dims 12:14).

Experts (4)	main h16 1x (16 rows) · n8 h8 1x (8 rows) · m8 h16 2.5x ([2,3,2,3,2,3] -> 6 rows / 15 steps) · m4 h8 2.5x ([2,3,3] -> 3 rows / 8 steps)
Merge rule	continuous dims summed over each block; discrete dims [6, 11] (gripper_close, control_mode) take the block's last value (--discrete-action-dims 6 11)
Gating	label-gated: conf head decides fine vs compressed group (pred >= tau), router picks horizon inside the group. Gate applies at inference only; all four experts are trained every step
conf target	conf * (1 - fixed) (VLM speed-ratio confidence, forced to 0 on contact frames)
conf head	scalar regression, MSE, conf_loss_coef=0.1, readout detached from state encoder
Router losses	soft_mixture + kl_supervise (0.1) + balance (0.05), router warmup 5,000 steps, min_prob 0.05
Expert init	m8 / m4 / n8 decoders initialised from 0.02*randn (default)
Training	60,000 steps, global batch 64 (32 x 2 GPU), AdamW lr 1e-4 cosine, warmup 0.05, seed 42, bf16, 2x H200 (Slurm job 16244, finished 2026-09-17)
Labels	prehj/robocasa-ratio-labels-contact
Code	rakybond007/GR00T-action-quantization branch jimin-dev-label-gated @ 3476549 (scripts/gr00t_finetune.py --use-moe-routing --moe-num-experts 4 --use-merged-8-head --use-merged-4-head --use-native-8-head --moe-label-gated --moe-speed 2.5 --discrete-action-dims 6 11 --data-config single_panda_gripper_conf)

Differs from easyminnn/atq-labelgated-2.5x-vlm_contact-60k only by --discrete-action-dims 6 11: that run summed the binary gripper / control_mode dims inside compressed blocks, this one takes the block's last value.

tau (conf_threshold) is an eval-time knob; the stored 0.5 is a default, not a trained constant. Output metadata carries _flevel_level / _flevel_k; the client executes the returned rows as-is (no re-merging).