Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer

Gemini Robotics Team, Abbas Abdolmaleki, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Ashwin Balakrishna, Nathan Batchelor, Alex Bewley, Jeff Bingham, Michael Bloesch, Konstantinos Bousmalis, Philemon Brakel, Anthony Brohan, Thomas Buschmann, Arunkumar Byravan, Serkan Cabi, Ken Caluwaerts, Federico Casarini, Christine Chan, Oscar Chang, London Chappellet-Volpini, Jose Enrique Chen, Xi Chen, Hao-Tien Lewis Chiang, Krzysztof Choromanski, Adrian Collister, David B. D'Ambrosio, Sudeep Dasari, Todor Davchev, Meet Kirankumar Dave, Coline Devin, Norman Di Palo, Tianli Ding, Carl Doersch, Adil Dostmohamed, Yilun Du, Debidatta Dwibedi, Sathish Thoppay Egambaram, Michael Elabd, Tom Erez, Xiaolin Fang, Claudio Fantacci, Cody Fong, Erik Frey, Chuyuan Fu, Ruiqi Gao, Marissa Giustina, Keerthana Gopalakrishnan, Laura Graesser, Oliver Groth, Agrim Gupta, Roland Hafner, Steven Hansen, Leonard Hasenclever, Sam Haves, Nicolas Heess, Brandon Hernaez, Alex Hofer, Jasmine Hsu, Lu Huang, Sandy H. Huang, Atil Iscen, Mithun George Jacob, Deepali Jain, Sally Jesmonth, Abhishek Jindal, Ryan Julian, Dmitry Kalashnikov, M. Emre Karagozler, Stefani Karp, Matija Kecman, J. Chase Kew, Donnie Kim, Frank Kim, Junkyung Kim, Thomas Kipf, Sean Kirmani, Ksenia Konyushkova, Li Yang Ku, Yuheng Kuang, Thomas Lampe, Antoine Laurens, Tuan Anh Le, Isabel Leal, Alex X. Lee, Tsang-Wei Edward Lee, Guy Lever, Jacky Liang, Li-Heng Lin, Fangchen Liu, Shangbang Long, Caden Lu, Sharath Maddineni, Anirudha Majumdar, Kevis-Kokitsi Maninis, Andrew Marmon, Sergio Martinez, Assaf Hurwitz Michaely, Niko Milonopoulos, Joss Moore, Robert Moreno, Michael Neunert, Francesco Nori, Joy Ortiz, Kenneth Oslund, Carolina Parada, Emilio Parisotto, Amaris Paryag, Acorn Pooley, Thomas Power, Alessio Quaglino, Haroon Qureshi, Rajkumar Vasudeva Raju, Helen Ran, Dushyant Rao, Kanishka Rao, Isaac Reid, David Rendleman, Krista Reymann, Miguel Rivas, Francesco Romano, Yulia Rubanova, Peter Pastor Sampedro, Pannag R Sanketi, Dhruv Shah, Mohit Sharma, Kathryn Shea, Mohit Shridhar, Charles Shu, Vikas Sindhwani, Sumeet Singh, Radu Soricut, Rachel Sterneck, Ian Storz, Razvan Surdulescu, Jie Tan, Jonathan Tompson, Saran Tunyasuvunakool, Jake Varley, Grace Vesom, Giulia Vezzani, Maria Bauza Villalonga, Oriol Vinyals, René Wagner, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Chengda Wu, Markus Wulfmeier, Fei Xia, Ted Xiao, Annie Xie, Jinyu Xie, Peng Xu, Sichun Xu, Ying Xu, Zhuo Xu, Jimmy Yan, Sherry Yang, Skye Yang, Yuxiang Yang, Hiu Hong Yu, Wenhao Yu, Wentao Yuan, Jingwei Zhang, Tingnan Zhang, Zhiyuan Zhang, Allan Zhou, Guangyao Zhou, Yuxiang Zhou

TL;DR

Google DeepMind has introduced Gemini Robotics 1.5, an agentic system combining a multi-embodiment Vision-Language-Action (VLA) model with a state-of-the-art Embodied Reasoning (ER) model. By introducing a novel Motion Transfer mechanism and an internal natural language "thinking" process, the system enables zero-shot skill transfer across radically different robot forms and dramatically improves performance on complex, long-horizon tasks.

What Problem Does This Paper Address?

Building truly general-purpose robots requires a system that can seamlessly perceive the environment, reason through complex tasks, and execute precise physical actions across diverse hardware.

Historically, robotic models have been siloed: a model trained for a specific bimanual arm setup couldn't easily transfer its skills to a humanoid robot without extensive post-training or explicit action-space alignment. Furthermore, executing long-horizon, multi-step tasks (like packing a suitcase based on a weather forecast) remains a massive hurdle. Robots typically struggle to break down high-level instructions into robust low-level actions, detect intermediate failures, or adaptively replan when things go wrong.

Methods

The authors address these challenges by combining two specialized models into a unified, agentic framework:

  1. Gemini Robotics-ER 1.5 (The Orchestrator): A Vision-Language Model (VLM) optimized for high-level embodied reasoning tasks like visual/spatial understanding, task planning, and progress estimation. It processes user inputs, leverages digital tools (like web search or code execution), and breaks complex tasks into simpler steps.
  2. Gemini Robotics 1.5 (The Action Model): A multi-embodiment VLA model that translates the orchestrator's instructions into low-level robot actions.
Overview of the Gemini Robotics 1.5 family of models and agentic framework.

Two core innovations power this iteration:

  • Motion Transfer (MT): A novel architecture and training recipe that allows the VLA model to learn from heterogeneous, multi-robot data sources. This forms a unified understanding of motion and physics, enabling zero-shot skill transfer between entirely different robot form factors.
  • Embodied Thinking (Thinking VLA): Similar to chain-of-thought reasoning in large language models, the VLA model is trained to "think before acting." It generates natural language "thinking traces" (e.g., decomposing a task or describing a motion trajectory) and appends them to its context window before emitting a physical action.

Datasets

The models were trained on a massive dataset combining internet-scale text, image, and video data with multi-embodiment robot data collected from thousands of diverse manipulation tasks across three platforms: ALOHA 2, Bi-arm Franka, and the Apptronik Apollo humanoid robot.

Evaluation was conducted using:

  • A benchmark of 230 tasks spanning multiple axes of variation (visual, instruction, action, and task generalization).
  • A newly designed Cross-Embodiment Benchmark to test zero-shot skill transfer between different physical platforms.
  • ASIMOV-2.0: An upgraded physical safety benchmark featuring real-world injury scenarios, video modalities, and physical constraint reasoning tasks.

Key Insights and Ideas

  • Thinking Traces Improve Physical Robustness: Interleaving natural language reasoning with physical movements allows the model to map high-level commands to low-level actions far more reliably than end-to-end translation. The visual-linguistic backbone handles the abstract planning, leaving a simpler action mapping for execution.
  • Spontaneous Error Recovery: Because the Thinking VLA continuously evaluates its state in natural language, it exhibits implicit situational awareness. If an object slips out of a robot's right hand, the model's next internal thought automatically updates to "pick up the object with the left hand," initiating self-correction without needing an external success detector.
  • Cross-Embodiment Synergy: Training on data from multiple distinct robot forms boosts the generalization of all embodiments. The Motion Transfer mechanism successfully bridges the gap between highly dissimilar forms (e.g., bimanual arms vs. humanoids) by extracting shared physical commonalities.

Key Results

  • Superior Generalization: Gemini Robotics 1.5 consistently outperformed previous baselines across all robots, showing substantial gains in task and action generalization.
  • Zero-Shot Skill Transfer: The system demonstrated successful zero-shot transfer; for example, the ALOHA robot successfully executed tasks for which training data was exclusively collected on the Bi-arm Franka platform.
  • Thinking Mode Boosts Progress: Enabling the internal reasoning process yielded a sizable improvement in progress scores for complex multi-step tasks across all platforms (e.g., increasing the ALOHA progress score from 0.26 to 0.55).
Task progress in the multi-step benchmark with and without enabling thinking during inference.
  • Halved Failure Rates: When evaluating long-horizon tasks, the combined GR-ER 1.5 + GR 1.5 agentic system cut total failure rates in half (from 44.5% to 22%) compared to an off-the-shelf Gemini 2.5 Flash orchestrator, primarily due to a sharp drop in planning errors (from 25.5% to 9%).

Limitations and Future Work

Despite major leaps in reasoning and generalization, the authors note that the model's low-level dexterity remains on par with the previous generation. The architectural changes lay the groundwork for generalist control, but performing highly intricate, precise manipulations remains a bottleneck.

To address data scarcity—especially for humanoids—future work will focus on scaling training data by learning directly from unannotated, low-quality public human videos and synthetic video corpora. Additionally, the team plans to explore reinforcement learning techniques to enhance dexterity without sacrificing the model's broad generality.

Summary updated

Abstract

General-purpose robots need a deep understanding of the physical world, advanced reasoning, and general and dexterous control. This report introduces the latest generation of the Gemini Robotics model family: Gemini Robotics 1.5, a multi-embodiment Vision-Language-Action (VLA) model, and Gemini Robotics-ER 1.5, a state-of-the-art Embodied Reasoning (ER) model. We are bringing together three major innovations. First, Gemini Robotics 1.5 features a novel architecture and a Motion Transfer (MT) mechanism, which enables it to learn from heterogeneous, multi-embodiment robot data and makes the VLA more general. Second, Gemini Robotics 1.5 interleaves actions with a multi-level internal reasoning process in natural language. This enables the robot to "think before acting" and notably improves its ability to decompose and execute complex, multi-step tasks, and also makes the robot's behavior more interpretable to the user. Third, Gemini Robotics-ER 1.5 establishes a new state-of-the-art for embodied reasoning, i.e., for reasoning capabilities that are critical for robots, such as visual and spatial understanding, task planning, and progress estimation. Together, this family of models takes us a step towards an era of physical agents-enabling robots to perceive, think and then act so they can solve complex multi-step tasks.
BibTeX
@misc{team2025gemini,
  author = {Gemini Robotics Team and Abbas Abdolmaleki and Saminda Abeyruwan and Joshua Ainslie and Jean-Baptiste Alayrac and Montserrat Gonzalez Arenas and Ashwin Balakrishna and Nathan Batchelor and Alex Bewley and Jeff Bingham and Michael Bloesch and Konstantinos Bousmalis and Philemon Brakel and Anthony Brohan and Thomas Buschmann and Arunkumar Byravan and Serkan Cabi and Ken Caluwaerts and Federico Casarini and Christine Chan and Oscar Chang and London Chappellet-Volpini and Jose Enrique Chen and Xi Chen and Hao-Tien Lewis Chiang and Krzysztof Choromanski and Adrian Collister and David B. D'Ambrosio and Sudeep Dasari and Todor Davchev and Meet Kirankumar Dave and Coline Devin and Norman Di Palo and Tianli Ding and Carl Doersch and Adil Dostmohamed and Yilun Du and Debidatta Dwibedi and Sathish Thoppay Egambaram and Michael Elabd and Tom Erez and Xiaolin Fang and Claudio Fantacci and Cody Fong and Erik Frey and Chuyuan Fu and Ruiqi Gao and Marissa Giustina and Keerthana Gopalakrishnan and Laura Graesser and Oliver Groth and Agrim Gupta and Roland Hafner and Steven Hansen and Leonard Hasenclever and Sam Haves and Nicolas Heess and Brandon Hernaez and Alex Hofer and Jasmine Hsu and Lu Huang and Sandy H. Huang and Atil Iscen and Mithun George Jacob and Deepali Jain and Sally Jesmonth and Abhishek Jindal and Ryan Julian and Dmitry Kalashnikov and M. Emre Karagozler and Stefani Karp and Matija Kecman and J. Chase Kew and Donnie Kim and Frank Kim and Junkyung Kim and Thomas Kipf and Sean Kirmani and Ksenia Konyushkova and Li Yang Ku and Yuheng Kuang and Thomas Lampe and Antoine Laurens and Tuan Anh Le and Isabel Leal and Alex X. Lee and Tsang-Wei Edward Lee and Guy Lever and Jacky Liang and Li-Heng Lin and Fangchen Liu and Shangbang Long and Caden Lu and Sharath Maddineni and Anirudha Majumdar and Kevis-Kokitsi Maninis and Andrew Marmon and Sergio Martinez and Assaf Hurwitz Michaely and Niko Milonopoulos and Joss Moore and Robert Moreno and Michael Neunert and Francesco Nori and Joy Ortiz and Kenneth Oslund and Carolina Parada and Emilio Parisotto and Amaris Paryag and Acorn Pooley and Thomas Power and Alessio Quaglino and Haroon Qureshi and Rajkumar Vasudeva Raju and Helen Ran and Dushyant Rao and Kanishka Rao and Isaac Reid and David Rendleman and Krista Reymann and Miguel Rivas and Francesco Romano and Yulia Rubanova and Peter Pastor Sampedro and Pannag R Sanketi and Dhruv Shah and Mohit Sharma and Kathryn Shea and Mohit Shridhar and Charles Shu and Vikas Sindhwani and Sumeet Singh and Radu Soricut and Rachel Sterneck and Ian Storz and Razvan Surdulescu and Jie Tan and Jonathan Tompson and Saran Tunyasuvunakool and Jake Varley and Grace Vesom and Giulia Vezzani and Maria Bauza Villalonga and Oriol Vinyals and René Wagner and Ayzaan Wahid and Stefan Welker and Paul Wohlhart and Chengda Wu and Markus Wulfmeier and Fei Xia and Ted Xiao and Annie Xie and Jinyu Xie and Peng Xu and Sichun Xu and Ying Xu and Zhuo Xu and Jimmy Yan and Sherry Yang and Skye Yang and Yuxiang Yang and Hiu Hong Yu and Wenhao Yu and Wentao Yuan and Yuan Yuan and Jingwei Zhang and Tingnan Zhang and Zhiyuan Zhang and Allan Zhou and Guangyao Zhou and Yuxiang Zhou},
  title = {{Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer}},
  year = {2025},
  eprint = {2510.03342},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.03342},
}
Figure