openvtla_insert_USB (baseline)

pi0.5 + AnyTouch2 (OpenVTLA-style re-implementation) trained on OpenViTac insert_USB (simulation, GelSight Mini, 100 demos) with tactile_base (https://github.com/YuriPark0123/tactile_base), config openvtla_insert_USB, exp baseline.

  • Code: tactile_base 831a50b (single-task training is unchanged in later commits)
  • Data: FVL-Lab/OpenViTac rev 5fb2757b88826917242adefd9196424ef8cf2d98, AnyTouch2 checkpoint-4frames features (GPU fp16)
  • Training: 20k steps, batch 64 (16 x 4 grad accum), seed 42, LoRA (gemma_2b_lora + gemma_300m_lora), cosine LR peak 5e-5, horizon 50, 18 tactile tokens (cls_pool2), 1x RTX 3090, ~4.6 days (2026-11-12 -> 11-17)
  • Final train loss ~0.0002 (see train_log.txt). Not yet evaluated in the simulator.

Contents: 19999/ = final openpi checkpoint (params/ for inference, train_state/ to resume, assets/ = norm stats). Place it at checkpoints/openvtla_insert_USB/baseline/19999 in a tactile_base checkout.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support