What Is Hume?

Hume is introduced in arXiv 2505.21432 by Haoming Song, Delin Qu, Yuanqi Yao, Qizhi Chen, Qi Lv and seven co-authors (v1 27 May 2025, v4 8 Jul 2025), with an official project page and a GitHub repository. The abstract says slow thinking, which has boosted LLMs on digital tasks, is largely unexplored for robotic foundation models. Read: the arXiv abstract page, project page and repository README; the full paper text was not read.

Third-party research; Firsthand is not affiliated and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.

The abstract describes Hume as a dual-system VLA. System 2 is the slow part and deliberates through value-guided thinking. System 1 is a lightweight reactive visuomotor policy that produces fluid actions. At deployment System 2 runs at a low frequency while System 1 receives its selected action asynchronously and predicts actions in real time.

System 2 extends a VLA backbone with a value-query head that estimates the state-action value of predicted actions. It repeatedly samples several candidate action chunks, scores each, and selects one by state-action value. The project page describes N candidates generated at different noise levels, with the best-of-N by highest Q value passed on as the optimal candidate. This is selection among actions; the sources read describe no written text plan.

System 1 takes the System-2-selected action and performs cascaded action denoising. The project page says the selected candidate is segmented and conveyed to System 1 for continuous action denoising. The README example sets s2_candidates_num to 5; that is an example setting, not a result.

The project page says 3 simulation environments and 3 real-world robotic platforms, covering 15 robot learning scenarios and 21 real-world manipulation tasks. Its galleries name a WidowX 250s, a Franka Emika Panda and an Agibot G-1, with tasks such as put carrot on plate, pour water and fold shorts. The abstract claims Hume outperforms existing state-of-the-art VLAs; no scores were read, so none are given here.

The page also shows failure-recovery clips where it says Pi0 and GR00T stay trapped in error and Hume recovers. That is the page’s claim, not verified here.

The hume-vla/hume repository states the MIT licence. Its README dates the training codebase to 3 June 2025 and the LIBERO evaluation code and checkpoints to 8 June 2025; real-world and Simpler evaluation are unchecked on its TODO list. The model zoo lists Hume-System2, pretrained weights exported from pi0, and the dataset list names LeRobot-format LIBERO, Bridge and Fractal sets under IPEC-COMMUNITY.

The site entries describe Hi Robot as a high-level vision-language model writing a language command for a low-level VLA, ECoT as a model writing a text plan, sub-task and grounded features before each action, and ThinkAct as a reasoning model passing a visual plan latent to an action model. Hume samples action candidates, selects one with a learned value, then denoises it.

Benchmark scores, training data counts, compute and model sizes were not read and are not given here. The project page and the README citation list anonymous authors, while the arXiv page names twelve; affiliations were not stated in the sources read. Whether first-person human video would suit Hume is Not stated.

The README example passes two camera images, a robot state vector and one task string per observation. Firsthand captures human first-person video, not robot data or robot state, and states one task instruction per episode; value labels and action candidates are Not stated. Write tasks, environments and signals into a spec and confirm on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats: RLDS, LeRobot v2.1, WebDataset, HDF5; whether a delivery matches Hume’s expected fields is the buyer’s to check.

What is Hume?

A dual-system VLA (arXiv 2505.21432): System 2 samples candidate actions and selects one with a value-query head; System 1 denoises it into real-time actions.

What is value-guided action selection?

Sampling several candidate action chunks and choosing the one with the highest estimated state-action value, as the abstract and project page describe.

What does Hume claim on benchmarks?

The abstract claims it outperforms existing state-of-the-art VLAs in simulation and on real robots. No scores were read, so none are given here.

Is Hume open source?

The hume-vla/hume README states the MIT licence, with parts under other sources’ licences. It releases training code, LIBERO evaluation and checkpoints; real-world evaluation is unchecked on its TODO list.

Does Firsthand use Hume, and can I request a custom dataset?

Firsthand is not affiliated and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.