What Is HybridVLA?

HybridVLA is introduced in arXiv 2503.10631, first submitted on 13 March 2025 and last revised on 23 June 2025 (v3) by Jiaming Liu, Hao Chen and 13 co-authors. The project page lists Peking University, The Chinese University of Hong Kong, BAAI and Fudan University. This page draws on the abstract, the official project page and the official GitHub README. The full paper text was not read, so anything beyond those sources is Not stated in the sources read.

The abstract says autoregressive VLA methods quantize actions into discrete bins, which disrupts the continuity needed for precise control. Existing diffusion-based VLA methods add a diffusion head that predicts continuous actions solely from VLM feature representations, without fully using the VLM’s pretrained reasoning through token-level generation. HybridVLA aims to absorb the continuity of diffusion actions and the contextual reasoning of autoregression within a single large language model.

Per the abstract, a collaborative training recipe incorporates diffusion denoising into the next-token prediction process to mitigate interference between the two paradigms, and a collaborative action ensemble mechanism adaptively fuses both predictions. The project page says inputs become tokens in the LLM’s embedding space, the denoising timestep and noise are projected to continuous vectors, and BOD and EOD marker tokens bridge the two paradigms. The README says the model is pretrained on large, diverse, cross-embodied real-world robotic datasets, then fine-tuned on simulation and self-collected real-world data, with code built on OpenVLA and CogACT. Parameter counts, dataset sizes and compute are Not stated in the sources read.

The arXiv v3 abstract says HybridVLA outperforms previous state-of-the-art VLA methods by 14% and 19% in mean success rate on simulation and real-world tasks, with stable manipulation in unseen configurations. The project page gives 17% and 19%, a mismatch this page does not resolve. The README says it evaluated in RLBench, built on the CoppeliaSim simulator, and shows generalization to unseen objects, backgrounds, spatial positions and lighting. Per-task scores and the baselines compared are Not stated in the sources read.

The README says HybridVLA’s code builds on OpenVLA and CogACT, and this site has a CogACT entry. A diffusion policy, as this site defines it, generates actions by iterative denoising; HybridVLA places that denoising inside an LLM’s token sequence instead of a separate head. InstructVLA targets lost vision-language ability, ChatVLA unifies understanding and control with separate experts, and OneTwoVLA switches between reasoning and acting.

HybridVLA’s data is robot data. Firsthand supplies real human egocentric demonstrations, never synthetic, generated or simulated, and the site describes no robot action, teleoperation, simulation, visual question answering or caption data. Its only language field is one task instruction per episode. Firsthand makes no claim that its data trained, was used with or is compatible with HybridVLA or any other model, robot or benchmark.

Firsthand collects real egocentric demonstrations on demand, to a written spec: a head-mounted global-shutter camera plus two wrist cameras, with 3D hand pose at 21 joints per hand with a per-joint visibility flag, wrist trajectories, contact events, action segments and one task instruction per episode. You confirm the spec on a Spec Pilot and pay per validated hour; the site states a 48-hour median turnaround where coverage is deep. Delivery is RLDS, LeRobot, WebDataset or HDF5. Request a custom dataset below.

What is HybridVLA?

A vision-language-action model introduced in arXiv 2503.10631. It predicts robot actions with both diffusion and autoregression inside one large language model, and fuses the two predictions at inference.

How does HybridVLA combine diffusion and autoregression?

Per the abstract, a collaborative training recipe incorporates diffusion denoising into the next-token prediction process, which mitigates interference between the two paradigms inside a single LLM.

What is the collaborative action ensemble?

Per the abstract, a mechanism that adaptively fuses the diffusion and autoregressive predictions for more robust control. The README exposes three modes: diff+ar, ar and diff.

What data does HybridVLA use?

The README says pretraining on large, diverse, cross-embodied real-world robotic datasets, then fine-tuning on simulation and self-collected real-world data. Dataset sizes are Not stated in the sources read.

Can HybridVLA be used commercially?

The README says the project is licensed under the MIT License. Terms for any dataset or checkpoint beyond that are Not stated in the sources read. Check the repository; this is not legal advice.

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

Firsthand is not affiliated with HybridVLA’s authors and makes no claim of use, training or compatibility. Firsthand data is real human egocentric capture. You can request a custom egocentric dataset: write a spec, confirm it on a Spec Pilot, pay per validated hour.