Hy-Embodied

Scroll sideways for more →

Hy-Embodied is a family of open-weight embodied-AI foundation models built by Tencent, jointly developed by the Tencent Hunyuan model team (the "Hy Team") and the Tencent Robotics X laboratory. Presented together as a matrix at the 2026 World Artificial Intelligence Conference (WAIC, held in Shanghai from July 17 to 20, 2026), the family splits a robot's software intelligence into three cooperating foundation models: Hy-Embodied-VLM-1.0 for perception, Hy-Embodied-RxBrain-1.0 for reasoning and planning, and Hy-Embodied-VLA-0.5 for action. All three are released under the Apache License 2.0 and form the model layer of Tencent's Tairos embodied-intelligence platform.

What Hy-Embodied is

Hy-Embodied is not a single network but a division of labor across the "perceive, think, act" loop that robotics researchers use to describe an embodied agent. Each model is a distinct release with its own code repository, model card, and technical report ^{[1][2][10][11]}:

• Hy-Embodied-VLM-1.0 is a vision-language model tuned for embodied perception: recognizing objects and their affordances, understanding 3D space, and reading how a scene changes over time ^[6].
• Hy-Embodied-RxBrain-1.0 is an embodied-cognition model that unifies language reasoning with visual imagination. It outputs both a plan (the steps) and a predicted picture of the goal state (what success should look like) in a single sequence ^[11].
• Hy-Embodied-VLA-0.5 is a vision-language-action model that converts a high-level goal into a stream of continuous, correctable robot motions, and ships with a full "data to model to training to deployment" learning stack ^[3][10].

The design intent, stated by Tencent Robotics X chief scientist Zhengyou Zhang at WAIC, is that "true intelligence emerges when language, vision, spatial understanding, physical control and environmental feedback work together" ^[1]. Rather than one monolithic policy, Hy-Embodied bets that perception, cognition, and control are best trained as specialized models that a runtime agent orchestrates.

A note on naming: Tencent's own repositories and model cards use both "Hy-Embodied" and "HY-Embodied," and mix version numbers (0.5 and 1.0) across components, so the capitalization and version tags are not internally consistent ^[2][3][5][7]. This article uses "Hy-Embodied" for the family and the exact tag (for example "VLA-0.5") for each component. The three headline models announced at WAIC are VLM-1.0, RxBrain-1.0, and VLA-0.5; the earlier HY-Embodied-0.5 and HY-Embodied-0.5-X releases from April 2026 are the lineage the VLM grew out of ^[4][7][9].

Background

Tencent Robotics X and the Hunyuan family

Tencent Robotics X is the robotics lab Tencent founded in 2018, led by Zhengyou Zhang, a computer-vision researcher known for the widely used Zhang camera-calibration method. The lab has produced quadruped and wheeled-legged research robots and a line of internal demonstration robots. Separately, "Hunyuan" is Tencent's foundation-model brand, spanning the Hunyuan large language models, image and 3D generators such as Hunyuan3D, and video and world generators such as HunyuanVideo and HunyuanWorld. Hy-Embodied is a collaboration between these two groups, and it should not be confused with those general-purpose Hunyuan models: it is a separate, robotics-specific line, and (unlike most Hunyuan LLM weights, which ship under the custom Tencent Hunyuan Community License) it is released under Apache 2.0 ^[1][6].

Why a platform company enters embodied AI

Tencent has been explicit that it does not intend to build robot bodies. In a WAIC 2026 interview headlined "why we do not build robot bodies, only brains," Zhang framed Tencent's role as supplying the software intelligence, models, tools, and data, that the hardware industry can plug in ^[16]. Tencent calls its platform Tairos, whose Chinese name 钛螺丝 ("titanium screw") casts the company as a small, reliable, universal component that holds the wider robotics industry together rather than a competitor to robot makers ^[17][18]. Hy-Embodied is the model layer of that strategy. Chinese coverage summarized the pitch as moving embodied AI out of the "brain in a vat," giving a capable model a body, sensors, and a feedback loop ^[13].

The three models

Scroll sideways for more →

Hy-Embodied-VLM-1.0 (perception)

VLM-1.0 is the family's eyes. It is a mixture-of-experts vision-language model of roughly 30B total parameters that activates only about 3B per token (8 of 128 experts plus one shared expert), pairing an Hy3-A3B language backbone with a native-resolution Hy-ViT2 vision encoder and a 32,768-token context window ^[6]. It is the production successor to the April 2026 HY-Embodied-0.5 base model, whose report described two variants, an efficient 2B-activated MoT and a 32B-activated model, evaluated across 22 benchmarks of visual perception, spatial reasoning, and embodied understanding ^[9].

Tencent groups VLM-1.0's target skills into three tiers, glossed poetically in Chinese as "see an object and know its use," "see a scene and know how it changes," and "see far and know the way back," which correspond to the model card's Action-Relevant State Understanding (23 benchmarks), Action-Transition Reasoning (8), and Sequential and Adaptive Reasoning (7) categories, 38 evaluations in total ^[6][12]. The card reports best-in-class-size performance on 19 of those 38 and second place on 11 more, and states that "despite activating only 3B parameters, it achieves performance close to that of the previous-generation model with 32B activated parameters" ^[6]. Tencent's press materials compress this into the headline that VLM-1.0 matches the previous flagship using about one-tenth of the compute (some Chinese coverage rounds the count to "37 tasks") ^[1][12]. These are vendor-reported numbers on Tencent's own evaluation suite.

Hy-Embodied-RxBrain-1.0 (cognition)

RxBrain is the reasoning engine, and the most conceptually novel of the three. Where a VLM mainly describes a scene and a world model mainly predicts future frames, RxBrain does both inside one autoregressive sequence: it writes out a plan in language (task decomposition, temporal ordering, constraints, decision logic) and, at points chosen by a learned control token, generates an imagined image of the intended intermediate or final state ^[11]. Language supplies the abstract structure of the plan; visual imagination grounds each step in a concrete predicted world state.

Architecturally it is a unified multimodal Mixture-of-Transformers of about 6.2B parameters with separate pathways for text, vision, and image generation, so understanding and image synthesis share one model instead of separate towers. Imagined frames are produced by a flow-matching head that decodes into a frozen FLUX variational-autoencoder latent space ^[11]. RxBrain was co-developed by Tencent Robotics X, Futian Laboratory, and the Hy Team, and the report shows it can be extended to continuous robot action generation with "promising real-robot performance without large-scale action-data pretraining" ^[11]. As of late July 2026 it was the least mature release: its repository listed the RxBrain-Bench benchmark and parts of the fine-tuning code as still to come ^[8][11].

Hy-Embodied-VLA-0.5 (action)

VLA-0.5 (abbreviated HyVLA-0.5) is the model that actually moves a robot. It is described not just as a network but as "an end-to-end system that spans the full robot learning stack: data collection, model design, continued pre-training and supervised fine-tuning, RL post-training, and real-world deployment" ^[10]. It takes a language goal plus camera views and emits chunks of continuous end-effector motion that can be corrected on the fly. It is built on the Hy-Embodied-0.5 MoT backbone and adds a 370M-parameter flow-matching action expert (described below), and it is designed to be deployed across different robot bodies rather than a single arm ^[3][10]. Its headline manipulation result, 90.9% success on the "clean" split and 90.1% on the "randomized" split of the RoboTwin 2.0 simulation benchmark (50 tasks), is discussed under results below ^[10].

Architecture in detail

Hy-Embodied's action model combines four ideas that are worth unpacking for a general reader, because each targets a specific weakness of earlier VLA models ^[3][10].

Mixture-of-Transformers (MoT) backbone. A standard transformer runs every token, whether it came from an image or from text, through the same weights. A Mixture-of-Transformers keeps separate parameter sets for each modality: visual tokens are routed through vision-specific weights and text tokens through language weights, and the two only interact inside shared self-attention layers. Because only the relevant weights fire for a given token, the compact HY-Embodied-0.5 MoT variant has about 4B total parameters but activates only about 2.2B at inference, giving it the speed of a dense 2B model while keeping richer, modality-specialized representations ^[7][9]. This is different from a mixture-of-experts model, which routes by token content rather than by modality.

Dual-tower flow-matching action expert. Robot actions are continuous, so Hy-VLA does not predict discrete action tokens the way RT-2 does. Instead it attaches a dedicated 370M-parameter "action expert" that models the continuous action distribution using conditional flow matching, a generative technique (a cousin of diffusion) that learns to transport noise into a realistic action sequence ^[3][10]. The VLM tower and the action-expert tower share attention layers, so the grounded vision-language context directly conditions the motion the expert generates. This dual-tower flow-matching design mirrors the approach Physical Intelligence took with pi-zero, and Hy-VLA reports substantial gains over it on RoboTwin 2.0.

Compact memory encoder. Single-frame policies forget what just happened; adding history usually costs parameters and compute. Hy-VLA instead folds a short window of multi-view history into the vision encoder using factorized temporal-spatial attention, causal attention across frames and bidirectional attention within a frame, that reuses the existing query, key, value, and output projections and adds a fixed sinusoidal time code. The result introduces no new learnable parameters relative to the single-image Hy-ViT 2.0 while still giving the policy a memory of the last several frames (K=6 during fine-tuning) ^[10].

Delta-chunk action representation. To transfer across robot bodies, Hy-VLA does not predict absolute joint angles, which differ per machine. Each future action is a "delta-chunk" expressed in the end-effector's own reference frame relative to the current state: a 3D Cartesian translation, a 6D rotation, and a 1D normalized gripper command, predicted in chunks of 50 steps at 10 Hz ^[10]. Because the representation is decoupled from any one robot's kinematics, a policy pre-trained on the handheld data-collection rig can be fine-tuned onto different arms. Tencent reports cross-embodiment transfer of 90% on a JAKA K1 and 89% on an Astribot S1 using only the handheld pre-training data ^[10].

The data pipeline

The most distinctive part of the VLA release is the data engine. Hy-Embodied's action model is trained on a corpus of more than 10,000 hours of human demonstrations collected with a custom, sub-millimeter-precision handheld rig in the style of the Universal Manipulation Interface (UMI). UMI, introduced by Cheng Chi and colleagues at RSS 2024, is a hand-held parallel-jaw gripper carrying a single wide-angle camera; a person uses it to demonstrate a task directly, and the recording, because it is expressed at the gripper rather than at a specific robot, yields hardware-agnostic training data ^[19]. Tencent's version replaces the consumer camera with optical motion capture to reach sub-millimeter accuracy in a single, globally consistent world frame, and captures three camera views at 224 by 320 resolution ^[10].

The full internal corpus is reported as more than 1 million episodes and over 10,000 hours across 70 distinct tasks; Tencent publicly released a subset of more than 2,000 hours and over 250,000 episodes in the Lance format, compatible with LeRobot v3.0 ^[3][10]. Measured in hours of demonstration, that corpus is large relative to the open robot-manipulation datasets it would be compared against, though the comparison is imperfect because these datasets differ in how they were collected (handheld human demonstration versus teleoperated robots) and in trajectory length.

Scroll sideways for more →

The headline is that Hy-Embodied's demonstration hours (10,000+) exceed the hour counts of DROID (350) and AgiBot World (about 2,976), while Open X-Embodiment leads on raw trajectory count. Hours are not a like-for-like measure of quality, but the volume of high-precision handheld data is the resource Tencent leans on most heavily in its materials ^[1][10].

Training and deployment stack

Hy-VLA is trained in stages that map onto the "data to model to training to deployment" learning stack the release advertises ^[10]:

1. Continued pre-training. The policy is trained on the full UMI corpus for about 200,000 steps at a global batch size of 1,024, learning general manipulation behavior decoupled from any specific robot.
2. Supervised fine-tuning (SFT). About 60,000 further steps adapt the policy to task-specific and robot-specific data, using a 6-frame history window and 50-step action chunks.
3. Reinforcement-learning post-training. An offline reinforcement-learning method the report calls FlowPRO applies preference optimization to the flow-matching policy, sharpening success on the target tasks beyond what imitation learning alone achieves.
4. Simulation. The related HY-Embodied-0.5-X agent model is integrated with a simulation framework called PlaygroundX, built on Tairos, in which the model produces a plan and revises it from environmental feedback through a "reason, execute, detect failure, replan" loop; it is reported to top 7 of 10 mainstream embodied task-planning benchmarks among edge-scale models and to be evaluated on 1,011 internal household tasks ^[4].
5. Real-world deployment. The same delta-chunk policy is deployed to physical arms, with reported real-world success on Dobot X-Trainer bimanual tasks ranging from 73% (zipping a pen case) to 94% (folding and storing glasses; inserting bottles) ^[3][10].

Reported results

On RoboTwin 2.0 (50 tasks, evaluated over many rollouts), Hy-VLA reports 90.9% on the clean split and 90.1% on the randomized split, against a pi-zero baseline of 65.9% and 58.4%, gains of 25.0 and 31.7 points respectively ^[10]. The following table collects the numbers Tencent published; all are vendor-reported.

Scroll sideways for more →

Relationship to Tairos and supported robot platforms

The X post that seeded coverage called Hy-Embodied "the brain of the Tairos platform." Tencent's own framing is more precise: Hy-Embodied is the model layer of Tairos, the modular embodied-intelligence open platform Tencent Robotics X and Futian Laboratory first launched in 2025 and upgraded at WAIC 2026 ^[1][17][18]. Tairos provides three things in a plug-and-play way, foundation models (Hy-Embodied), development tools, and data services, plus an agent runtime. At WAIC, Tencent added the TairosAgent framework and a continuously running agent called Apexio, whose three-layer design maps onto the three models: an upper cognition layer (RxBrain, activated on demand for deep reasoning), a middle perception-action layer (VLM, running continuous multimodal sensing at roughly 15 Hz), and a lower execution layer for higher-frequency reflex-like responses ^[12]. Tencent Cloud also announced an Embodied-AI-as-a-Service (EaaS) offering to deliver these capabilities from the cloud ^[1]. A dedicated article covers Tairos in depth.

Tencent reports that over the preceding year it worked with robot makers including Unitree, Dobot (Yuejiang), Leju, and Songyan Power to bring the platform onto different bodies, and Chinese coverage lists early Tairos partners spanning Unitree, AgiBot (Zhiyuan), Dobot, Leju, PaXini, Keenon, and EngineAI ^[12][14][17]. Because Hy-VLA's delta-chunk actions are decoupled from any one robot's kinematics, the same policy can in principle be fine-tuned onto these varied platforms.

The Xiaoliu demonstration robot at WAIC 2026

Tencent's booth centerpiece at WAIC 2026 was a robot masseur that became one of the show's viral attractions, drawing queues of visitors for a shoulder rub ^[14]. The robot is 小六 (Xiaoliu, "Little Six"), a Tencent Robotics X demonstration platform first shown publicly at WAIC 2026 as an iteration of the earlier 小五 (Xiaowu, "Little Five") ^[15]. (Some English-language notes rendered the name as "Xiaoliu / Little Willow"; the Chinese primary sources consistently write 小六, "Little Six.")

Xiaoliu uses cable-driven actuation for safe, compliant human contact, and a self-developed multimodal sensing system that fuses force, tactile, and visual signals ^[15]. Zhang Zhengyou described the sensing latencies as roughly 1 to 2 milliseconds for touch and 15 to 30 milliseconds for vision, the tight loop a machine needs to press on a person safely ^[14]. Working with a traditional-Chinese-medicine practitioner, the team collected the doctor's hand techniques and applied forces and used reinforcement learning to reproduce the massage trajectories and pressures; Xiaoliu performs several synchronized dual-arm techniques and adjusts force when it senses muscle stiffness or a person's tolerance ^[14][15]. Zhang said the team hopes to place Xiaoliu in real settings such as nursing homes to gather more data and keep iterating ^[14].

Two cautions on sourcing. Xiaoliu appears only in Chinese coverage; Tencent's own English press release on the WAIC solution does not mention it ^[1]. And Xiaoliu is a demonstration platform for the embodied stack, not itself a product built on a specific published Hy-Embodied checkpoint; Tencent describes the technique-learning as reinforcement learning over its sensing data rather than crediting a named model, so the robot is best understood as a showcase of the same "sense, reason, act" philosophy rather than a deployed VLA benchmark.

How it compares to other open VLA models

Hy-Embodied enters a field of robot foundation models that has grown quickly since 2023. The table below places the VLA member against the best-known recent policies. "Openness" is judged by whether weights are downloadable; several widely cited models are not.

Scroll sideways for more →

Several observations follow. Hy-VLA's closest design relative is pi-zero: both attach a flow-matching action expert to a pre-trained VLM and share attention between the language and action towers. Hy-VLA differs in its handheld data engine, its zero-parameter memory encoder, and its embodiment-decoupled delta-chunk actions. Among the truly open models, Hy-VLA, OpenVLA, pi-zero, and GR00T N1 all release weights, whereas RT-2 and Helix are described in papers and blog posts but not distributed. On the perception and cognition side, VLM-1.0 and RxBrain-1.0 have fewer direct open peers, since most competing systems fold perception and planning into a single policy rather than shipping them as separately downloadable models.

Significance and reception

Hy-Embodied is notable less for any single benchmark than for what it packages together. Three points stand out.

First, the release is unusually open for a major Chinese platform company. Tencent shipped three embodied foundation models, plus a 2,000-hour-plus slice of hard-to-collect handheld data, under Apache 2.0, a permissive OSI-approved license, at a time when many rivals keep robot policies behind partner agreements ^[1][3][6]. That the models are Apache-licensed rather than governed by the restrictive Tencent Hunyuan Community License used for Tencent's LLMs is itself a signal about how the company is trying to seed an ecosystem.

Second, the split into perception, cognition, and action models is a design statement. It contrasts with the single-policy trend of OpenVLA or pi-zero and instead resembles the dual-system (fast controller plus slow planner) architectures of GR00T N1 and Helix, extended to three tiers and made independently reusable through Tairos ^[1][23][24]. RxBrain's "reason in text, imagine the goal in pixels" formulation is the most research-forward piece and gives the family a planning story distinct from pure world models ^[11].

Third, the WAIC framing positions Tencent as infrastructure. By publicly declining to build robot bodies and instead offering models, tools, and data as the "titanium screw" of the industry, Tencent is making a platform play analogous to what it has done in cloud and gaming, aimed at the fast-growing Chinese humanoid and manipulation market ^[16][17]. Chinese coverage of the three-model launch was broadly positive on the openness and on Xiaoliu's crowd appeal ^[14][28], while a strand of WAIC commentary warned that the sector risks homogenization once the demos stop and real orders have to be filled ^[14].

Limitations and open questions

Most of Hy-Embodied's quantitative claims are vendor-reported on Tencent's own suites, including RoboTwin 2.0 and the 38-benchmark VLM evaluation; independent reproduction on standard external leaderboards was not yet available at release ^[6][10]. The "one-tenth the compute" line compares VLM-1.0 to Tencent's own previous model, not to third-party baselines ^[6]. RxBrain-1.0 was the least complete of the three, with its benchmark (RxBrain-Bench) and parts of its code listed as forthcoming, so its planning claims rest largely on the technical report ^[8][11]. The impressive demonstration-hour count (10,000+) is handheld human data, which is cheaper to gather but must still cross the gap to reliable robot execution, and only a fifth of it was released publicly ^[3][10]. Cross-embodiment success was shown on a handful of arms (JAKA K1, Astribot S1, Dobot X-Trainer), not across the full breadth of Tairos partners ^[10]. Finally, Xiaoliu, the most memorable public artifact, is a curated demonstration sourced only to Chinese media and is not tied to a specific released checkpoint, so it should be read as a showcase, not a validated capability ^[1][14].

ELI5

Imagine a robot that needs three kinds of brain. One brain is good at looking: it sees a cup on a table and understands what the cup is for and where it is in space. That is VLM-1.0. A second brain is good at thinking ahead: it makes a step-by-step plan and even pictures in its mind what the finished job should look like, then follows both. That is RxBrain-1.0. The third brain is good at moving: it turns "pick up the cup" into a smooth stream of hand motions and fixes its own mistakes as it goes. That is VLA-0.5. To teach the moving-brain, Tencent had people hold a special gripper and do thousands of hours of chores, recording every tiny motion, then let the robot learn from that. Tencent gives all three brains away for free so any robot company can use them, and it showed them off at a big AI fair in Shanghai with a robot named Little Six that gives shoulder massages.

See also

• Tairos
• Vision-language-action model
• Universal Manipulation Interface
• pi-zero
• Isaac GR00T
• Embodied AI
• Hunyuan
• Flow matching

References

1. Tencent, "Tencent Unveils Full-Stack Embodied Intelligence Solution; ADP 4.0 Launches Globally at WAIC 2026," Tencent Newsroom, July 2026. https://www.tencent.com/tencent-unveils-full-stack-embodied-intelligence-solution-at-waic-2026/ ↩
2. Tencent Hunyuan, "HY-Embodied: Embodied Foundation Models for Real-World Agents," GitHub. https://github.com/Tencent-Hunyuan/HY-Embodied ↩
3. Tencent Hunyuan, "Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack," GitHub. https://github.com/Tencent-Hunyuan/Hy-Embodied-0.5-VLA ↩
4. Tencent Hunyuan, "HY-Embodied-0.5-X: An Enhanced Embodied Foundation Model for Real-World Agents," GitHub. https://github.com/Tencent-Hunyuan/HY-Embodied-0.5-X ↩
5. Tencent Hunyuan, GitHub organization. https://github.com/tencent-hunyuan ↩
6. Tencent, "Hy-Embodied-VLM-1.0," Hugging Face model card, 2026-07-15. https://huggingface.co/tencent/Hy-Embodied-VLM-1.0 ↩
7. Tencent, "HY-Embodied-0.5," Hugging Face model card, 2026-04-09. https://huggingface.co/tencent/HY-Embodied-0.5 ↩
8. Tencent, "Hy-Embodied-RxBrain-1.0," Hugging Face model card, July 2026. https://huggingface.co/tencent/Hy-Embodied-RxBrain-1.0 ↩
9. Xumin Yu, Zuyan Liu, Ziyi Wang, et al. (Tencent Robotics X and HY Vision Team), "HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents," arXiv:2604.07430, 2026-04-08. https://arxiv.org/abs/2604.07430 ↩
10. He Zhang, Lingzhu Xiang, Haitao Lin, et al., "Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack," arXiv:2606.14409, 2026-06 (v2 2026-07-20). https://arxiv.org/abs/2606.14409 ↩
11. Haotian Liang, Mingkang Chen, Yufei Huang, et al., "RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination," arXiv:2607.14187, 2026-07-15. https://arxiv.org/abs/2607.14187 ↩
12. Sina Finance, "直击WAIC丨腾讯发布多款具身智能基座模型与智能体," 2026-07-19. https://finance.sina.com.cn/hy/hyjz/2026-07-19/doc-iniiimtc7924002.shtml ↩
13. Sina Finance, "腾讯WAIC亮出AI全栈答卷：具身智能破'缸中之脑'，ADP 4.0海外版、WorkBuddy APP版同步出海," 2026-07-19. https://finance.sina.com.cn/stock/relnews/hk/2026-07-19/doc-iniiirzc3487525.shtml ↩
14. NetEase (163.com), "腾讯机器人按摩师走红WAIC：'要送进养老院'," 2026-07. https://www.163.com/dy/article/L25FK4K105129QAF.html ↩
15. Sina Tech, "腾讯 Robotics X 实验室新一代示范性机器人'小六'首次公开亮相，系'小五'迭代之作," 2026-07-17. https://finance.sina.com.cn/tech/digi/2026-07-17/doc-iniickay1505497.shtml ↩
16. NetEase (163.com), "WAIC2026访谈 | 腾讯张正友：我们为什么不造机器人本体，只造大脑," 2026-07. https://www.163.com/dy/article/L2AC028D0550WHYR.html ↩
17. Tencent Cloud Developer Community, "腾讯发布国内首个Robotics X具身智能开放平台Tairos，打造机器人领域的'钛螺丝'," 2025. https://cloud.tencent.com/developer/news/2805038 ↩
18. Economic Information Daily (Xinhua), "腾讯联合福田实验室发布具身智能开放平台Tairos," 2025-07-28. http://jjckb.xinhuanet.com/20250728/ea7fd8dbba184e32884c5dafc10a766a/c.html ↩
19. Cheng Chi, Zhenjia Xu, Chuer Pan, et al., "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots," Robotics: Science and Systems (RSS) 2024, arXiv:2402.10329. https://arxiv.org/abs/2402.10329 ↩
20. Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, et al., "OpenVLA: An Open-Source Vision-Language-Action Model," arXiv:2406.09246, 2024. https://arxiv.org/abs/2406.09246 ↩
21. Physical Intelligence, "π0: A Vision-Language-Action Flow Model for General Robot Control," arXiv:2410.24164, 2024; "Open Sourcing π0," physical-intelligence.company, 2025-02-04. https://arxiv.org/abs/2410.24164 ↩
22. Anthony Brohan, Noah Brown, Justice Carbajal, et al. (Google DeepMind), "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control," arXiv:2307.15818, 2023. https://arxiv.org/abs/2307.15818 ↩
23. NVIDIA, "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots," arXiv:2503.14734, 2025. https://arxiv.org/abs/2503.14734 ↩
24. Figure AI, "Helix: A Vision-Language-Action Model for Generalist Humanoid Control," figure.ai, February 2025. https://www.figure.ai/news/helix ↩
25. AgiBot and Shanghai AI Lab (OpenDriveLab), "AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems," arXiv:2503.06669, 2025. https://arxiv.org/abs/2503.06669 ↩
26. Open X-Embodiment Collaboration, "Open X-Embodiment: Robotic Learning Datasets and RT-X Models," arXiv:2310.08864, 2023. https://arxiv.org/abs/2310.08864 ↩
27. Alexander Khazatsky, Karl Pertsch, Suraj Nair, et al., "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset," arXiv:2403.12945, 2024. https://arxiv.org/abs/2403.12945 ↩
28. Huacheng (Guangzhou Media Cluster), "让具身智能拥有大脑，腾讯一次发布三大具身基座模型丨WAIC2026," 2026-07-18. https://huacheng.gz-cmc.com/pages/2026/07/18/6d0b3a29d1f9419a87ae0cd58687a8c7.html ↩