supernon/fast-thinkact-checkpoints

Fast-ThinkAct reproduction checkpoints

Checkpoints from an independent, small-scale reproduction of Fast-ThinkAct with RoboTwin demonstrations. Source and setup: https://github.com/supernoooo/fast-thinkact . This is a research snapshot, not an official release of the paper authors.

Directory	Contents and intended use
policy_ftc	RDT action policy conditioned on the coordinate-CoT latent student's visual plan. Use with vlm_student.
policy_rdt	Plain RDT baseline action policy.
vlm_student	Coordinate-CoT latent student, waypoint head and verbalizer. Matching student for policy_ftc.
vlm_student_nl	Separate natural-language-CoT latent student experiment. Not the student used to train the released FTC policy.
vlm_teacher	Coordinate-CoT GRPO reasoning teacher.
vlm_teacher_nl	Natural-language-CoT GRPO reasoning teacher.
vlm_cot_sft_nl	Natural-language CoT supervised-finetuning checkpoint.

Download from the source repository root:

pip install 'huggingface-hub>=0.36,<1'
hf download supernon/fast-thinkact-checkpoints --local-dir checkpoints

Download only the matching inference pair and baseline:

hf download supernon/fast-thinkact-checkpoints --local-dir checkpoints \
  --include 'vlm_student/*' 'policy_ftc/*' 'policy_rdt/*'

Keep directory names unchanged. Student loaders require the Hugging Face VLM files plus student_heads.pt; verbalizer.pt supports verbalization experiments. Policy files contain PyTorch state dictionaries. The source repository contains the custom model definitions and deployment adapter.

Base SigLIP/T5 encoders, original base weights for training from scratch, RoboTwin assets and demonstration datasets are not included. Obtain those separately from their publishers. Upstream models and simulator components retain their respective licenses and usage conditions.

The training/evaluation setup is substantially smaller than the paper's. The reported trend for adjust_bottle used policy_ftc with vlm_student; it does not establish generalization across all tasks. See the source experiment notes for completed experiments, incomplete ablations and limitations. Full training and simulator evaluation require CUDA GPUs and an appropriately configured RoboTwin environment.

checkpoint_manifest.json lists all uploaded files and expected sizes. Credentials, caches, local paths, optimizer state and datasets are excluded.