joon-stack/univla-simpler-success50-h369-concatmh-hyp-pool-step5000

Full merged UniVLA checkpoint for Simpler success50 H369 concat-MH fine-tuning.

model: h369_concat_hyper_full
horizon style: h369_concat_mh
horizons: 3,6,9
decoder type: pool
step: 5000
latent action token length: 15
per-horizon latent action token length: 5
predicted action horizon: 10

This repository contains full merged VLA weights plus:

action_decoder-5000.pt
concat_mh_action_decoder-pool-5000.pt

Eval must use the concat-MH decoder implementation. The policy still consumes the current image and instruction; it does not require separate h3/h6/h9 images at eval time.

Downloads last month: 17

Safetensors

Model size

8B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support