PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Yuan Yuan, Jingwei Zhang, Tingnan Zhang, Zhiyuan Zhang, Allan Zhou, Guangyao Zhou, Yuxiang Zhou

arXiv:2510.03342v3cs.RO

TL;DR

Truly general robots require deep understanding of the physical world. This paper combines a multi-embodiment action model with embodied reasoning and internal thinking; the authors report improved handling of complex multi-step tasks and state-of-the-art embodied reasoning.

  • Problem

    Truly general robots require deep understanding of the physical world.

    Source: §1 · p. 1

  • Method

    The paper combines a general Vision-Language-Action model with embodied reasoning, heterogeneous robot-data learning, and internal reasoning interleaved with actions.

    Source: §7 · p. 22

  • Results

    Gemini Robotics-ER 1.5 outperforms frontier models across broad embodied-intelligence benchmarks, while thinking notably improves the VLA’s handling of complex multi-step tasks.

    Source: §1 · p. 2, §7 · p. 22

  • Takeaways & Limitations

    The authors present the combined models as a path toward general-purpose robots that decompose long-horizon tasks and translate reasoning into successful execution.

    Source: §7 · p. 22

    Gemini Robotics 1.5’s dexterity remains on par with the previous generation.

    Source: §7 · p. 22

Abstract

from arXiv · show

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.

1. Introduction

Gemini Robotics 1.5 introduces a model family pairing a multi-embodiment Vision-Language-Action model with an embodied-reasoning model, which can be combined in an agentic system. The family brings reasoning, planning, tool use, and low-level action together to help robots solve complex tasks.

1. Introduction

  • The GR 1.5 family pairs a multi-embodiment VLA with an embodied-reasoning VLM and combines them in an agentic system for complex problems.The system orchestrates user dialogue, high-level reasoning and planning, agentic tool use, and low-level action.Source: §1 · p. 1
  • A novel architecture and Motion Transfer mechanism let GR 1.5 learn across diverse robot data, control ALOHA, Bi-arm Franka, and Apollo without robot-specific post-training, and transfer skills zero-shot between robots.Multi-embodiment pre-training forms a unified understanding of motion and physics.Source: §1 · p. 1
  • As a Thinking VLA, GR 1.5 interleaves natural-language thoughts with movement to interpret observations, simplify instructions, detect task success or failure, propose recovery, and make actions more interpretable.Source: §1 · p. 2
  • GR-ER 1.5 advances embodied reasoning across visuo-spatial-temporal understanding and significantly outperforms other frontier models on a broad suite of benchmarks while retaining frontier-model capabilities and running considerably faster.Its physical understanding combines with tool use, video and audio communication, and code-writing capabilities for a broad range of applications.Source: §1 · p. 2
  • The agentic framework supports long-horizon execution through complex planning and adaptive orchestration, multimodal interaction, user-specified tools such as web search, and layered safety reasoning.Source: §1 · p. 2

2. Method Overview

Gemini Robotics 1.5 pairs a reasoning orchestrator with an action model that can think before acting, while Motion Transfer enables learning across robot embodiments. Its development uses diverse robot data and a mix of real-robot and simulation evaluation.

2.1. Model & Architecture

  • The agentic system uses GR-ER 1.5 as an orchestrator to plan tasks and detect success, and GR 1.5 as an action model that converts instructions into low-level actions.The orchestrator can use digital tools and sends open-vocabulary natural-language instructions to the action model.Source: §2.1 · p. 3
  • Embodied Thinking lets GR-ER 1.5 plan and update tasks, while the Thinking VLA generates natural-language reasoning traces before actions to compose primitive skills for multi-step tasks.The traces are appended to the context window, making the VLA’s reasoning more transparent in human-robot interactions.Source: §2.1 · pp. 3–4
  • Motion Transfer combines a new VLA architecture and training recipe to learn across robots and data sources, transferring skills across different embodiments.The approach aims to unify the model’s understanding of motions and physical interactions.Source: §2.1 · p. 4

2.2. Robot Data

  • Training data combines tasks from ALOHA, Bi-arm Franka, and Apollo humanoid robots with publicly available Internet text, image, and video datasets.The robot data spans thousands of diverse tasks and manipulation skills across many scenes.Source: §2.2 · p. 4

2.3. Evaluation

  • Over 90% of development evaluation episodes ran in simulation, enabling broader iteration while real-world evaluation remained necessary to determine model quality.Evaluations use interleaved A/B/n tests on real robots; MuJoCo scenes aligned to real scenes showed strong rank consistency, reducing the volume of hardware tests.Source: §2.3 · p. 4

3. Gemini Robotics 1.5 is a general multi-embodiment Vision-Language-Action Model

Gemini Robotics 1.5 uses one model across different robot embodiments and tasks, with reported out-of-the-box control, cross-embodiment skill transfer, and benefits from interleaving thinking with action.

3. Gemini Robotics 1.5 is a general multi-embodiment Vision-Language-Action Model

  • GR 1.5 controls ALOHA, Bi-arm Franka, and Apollo robots with the same checkpoint on varied tasks, without post-training for a specific embodiment or task.Figure 2 illustrates examples of this out-of-the-box capability.Source: §3 · p. 4, Figure 2 · p. 5

3.1. Gemini Robotics 1.5 can generalize to new environments and tasks

  • GR 1.5 outperforms the compared baselines across generalization categories, with substantial instruction, action, and task gains on ALOHA and gains across categories on Bi-arm Franka and Apollo.The Bi-arm Franka and Apollo comparisons are against GRoD checkpoints trained with less data; those models also require separate embodiment-specific checkpoints.Source: §3.1 · pp. 6–7, Figure 3 · p. 6
  • Adding data from other embodiments generally improves generalization, and the Motion Transfer recipe amplifies this benefit.The ablation compares single-embodiment and all-embodiment data, both without MT, against the MT training recipe.Source: §3.1 · pp. 7–8, Figure 4 · p. 7

3.2. Learning across different robot embodiments

  • GR 1.5 demonstrates zero-shot skill transfer across ALOHA, Bi-arm Franka, and humanoid robots, including skills seen only in another embodiment’s training data.In the cross-embodiment benchmark, single-embodiment models perform poorly, while cross-embodiment data with MT yields significantly better performance; reported success rates indicate completed tasks, not only partial progress.Source: §3.2 · pp. 8–9, Figure 5 · p. 8
  • The benefits of cross-embodiment data and MT depend on platform data quantity and embodiment gap: MT helps align shared knowledge, but is less pronounced for data-scarce humanoids.Adding external data provides the greatest boost for humanoids, while MT alignment may be less effective across the larger humanoid-to-other-robot gap.Source: §3.2 · pp. 8–9

3.3. Thinking Helps Acting

  • Enabling thinking improves progress on multi-step tasks by decomposing high-level instructions into short-horizon language steps before mapping them to actions.The language trace uses the VLM backbone’s visual-linguistic capabilities, followed by a simpler action-mapping stage.Source: §3.3 · p. 9, Figure 6 · p. 9
  • Thinking traces make planned actions inspectable, reflect task progress by shifting objectives after subtasks succeed, and enable self-correction when execution goes wrong.The examples include switching from grasping a tennis ball to placing it in a bag, and recovering a slipped bottle with the other hand.Source: §3.3 · pp. 9–10, Figure 7 · p. 10

4. Gemini Robotics-ER 1.5 is a generalist embodied reasoning model

Gemini Robotics-ER 1.5 is a Gemini-based multimodal model for grounded embodied reasoning, designed to provide high-level intelligence alongside a general VLA. It combines broad frontier-model capabilities with strong spatial reasoning and benefits from additional inference-time thinking.

4. Gemini Robotics-ER 1.5 is a generalist embodied reasoning model

  • GR-ER 1.5 targets grounded spatial and temporal reasoning, intuitive physics, causality, and affordances, and can provide high-level intelligence when combined with a general VLA.The report presents it as a multimodal thinking model based on Gemini for embodied reasoning.Source: §4 · pp. 10–11

4.1. Generality

  • Across 15 academic embodied-reasoning benchmarks, GR-ER 1.5 expands the generality–embodied-reasoning Pareto frontier and achieves state-of-the-art embodied reasoning while retaining broad frontier-model capabilities.The evaluation pairs embodied-reasoning benchmarks with a generality mix covering image understanding, science, and coding.Source: Figure 8 · p. 11, §4.1 · p. 11

4.2. Frontier capabilities for Embodied Reasoning

  • GR-ER 1.5 handles progress estimation across embodiments, including task-completion percentage, success detection, and video-frame unshuffling, with strong success-detection performance across real-time/offline and single-/multiview settings.Real-time evaluations simulate 5Hz inference with latency, while offline evaluations allow unlimited inference time.Source: Figure 12 · p. 14, Figure 13 · p. 14, §4.2 · p. 15
  • On a benchmark drawn from early testers’ real-world applications, GR-ER 1.5 outperforms Gemini Robotics-ER and contemporary state-of-the-art multimodal models on spatial understanding tasks such as detection and pointing.The examples include parsing an inventory-shelf image into a table presented as an HTML page.Source: Figure 14 · p. 15, §4.2 · p. 15

4.3. Thinking

  • Its thinking traces show the model identifying key image features before fine details, reasoning methodically, pointing while thinking, and performing relevant mathematical operations.Examples include reading a gauge and matching socks.Source: §4.3 · p. 16, Figure 15 · p. 16

6. Responsible Development and Safety

The section describes a layered safety approach for Gemini Robotics and reports findings on semantic safety, human-robot dialogue, and automated adversarial testing. It also emphasizes ongoing operational safety and monitoring societal impacts as part of responsible deployment.

Safety Approach

  • The safety approach combines semantic reasoning, respectful dialogue, pre-action safety thinking, and low-level physical safeguards such as collision avoidance.The report also says it continues implementing operational-safety practices codified in existing standards (International Organization for Standardization (ISO), 2016, 2025).Source: §6 · p. 19

Safe Human-Robot Dialog

  • Adversarial testing found GR-ER 1.5 strongly compliant with Gemini Safety policies (Comanici et al., 2025), supporting socially appropriate speech, gestures, and actions.The policies are designed to prevent harmful content such as hate speech, sexual explicitness, and disclosure of personally identifiable information.Source: §6 · pp. 19–20

Semantic Action Safety

  • GR-ER 1.5 improves risk recognition, action-consequence understanding, and intervention decisions over GR-ER, while Thinking about Safety fine-tuning achieves state-of-the-art safety-constrained pointing results.ASIMOV-2.0 adds improved benchmark quality, broader coverage of real-world injury scenarios, video modalities, and embodied reasoning tasks about physical constraints.Source: §6 · p. 20, Figure 19 · p. 21

Auto-Red-Teaming Framework

  • Auto-red-teaming found GR-ER 1.5, especially with Thinking, more robust to instruction obfuscation, hallucination elicitation, and content-safety attacks; AutoRaters and generated training data help mitigate vulnerabilities.The framework pits an Attacker against a Target robot model and uses an AutoRater to judge responses; attacks can target instructions, images, or rollout environments.Source: §6 · pp. 20–21

Responsible Deployment

  • The report treats societal impacts, alongside safety risks, as requiring proactive management, monitoring, responsible deployment, and transparent reporting.It directs readers to Appendix A for the Gemini Robotics model card.Source: §6 · pp. 21–22

7. Discussion … B. Gemini Robotics 1.5 is a general multi-embodiment Vision-Language-Action Model

Gemini Robotics 1.5 combines a reasoning-capable VLA, cross-embodiment learning, and embodied reasoning to advance general-purpose robotics, while the report identifies data scarcity and dexterity as continuing challenges.

7. Discussion

  • Thinking before acting, heterogeneous cross-embodiment learning, and state-of-the-art embodied reasoning are the report’s three contributions toward general-purpose robots.The report says internal monologue improves complex multi-step tasks, cross-platform learning supports skill transfer amid data scarcity, and embodied reasoning covers visual and spatial thinking, planning, progress estimation, and success detection.Source: §7 · p. 22
  • Future work will use human and synthetic videos without action annotations and explore reinforcement learning to improve dexterity without sacrificing generality.The report notes that GR 1.5’s dexterity remains on par with the previous generation and proposes learning from publicly available low-quality video corpora, among other sources.Source: §7 · p. 22

A. Model Card

Model summary

  • Gemini Robotics-ER 1.5 is a vision-language model, while Gemini Robotics 1.5 is a vision-language-action model for manipulation across tasks, scenes, and robots.Both accept text and images; ER 1.5 generates text, while Robotics 1.5 generates continuous robot-action values and can additionally generate text in thinking mode.Source: Model summary · p. 30

Model Data

  • Both models were trained on image, text, robot-sensor, and action data, with filtering for sensitive information and synthetic image captions.Synthetic captions were generated using Gemini and FlexCap (Dwibedi et al., 2024) and paired with original captions to help the model learn image details.Source: Model Data · p. 30

Implementation Frameworks

Evaluation

  • The model card directs readers to Sections 3, 4, and 6 for Gemini Robotics 1.5, Gemini Robotics-ER 1.5, and safety evaluation procedures and results.The supplied passages refer to those sections and do not state the evaluation outcomes.Source: Evaluation · pp. 30–31

B. Gemini Robotics 1.5 is a general multi-embodiment Vision-Language-Action Model

B.1. Rank consistency between evaluations in simulation and on real robots … B.2.1.1 New action generalization progress score definition

The appendix reports rank consistency between simulation and real-robot evaluations and describes an expanded generalization benchmark, including tasks that test composition of learned motions. It also specifies how the new action-generalization tasks are scored.

B.1. Rank consistency between evaluations in simulation and on real robots

  • Physics-based simulators scale up evaluations of Gemini Robotics 1.5 on new objects, scenes, and environments.Source: §B.1 · p. 32
  • Across tasks, success-rate rankings are consistent between simulation and real-robot evaluations, supporting rapid iteration with fewer slow real-robot tests.Figure 21 describes A/B tests in simulation and on real robots across a range of tasks.Source: Figure 21 · p. 32, §B.1 · p. 32

B.2. Generalization benchmark

B.2.1. ALOHA robot

  • In-distribution performance uses the 20-task dexterity benchmark, while progress-score definitions for in-distribution, visual, semantic, and action generalization come from prior work (Gemini-Robotics-Team et al., 2025).Source: §B.2.1 · p. 32

B.2.1.1 New action generalization progress score definition

  • The expanded action-generalization benchmark tests composing learned motions in novel ways, beyond prior tests of new object locations and shape variants (Gemini-Robotics-Team et al., 2025).Source: §B.2.1.1 · p. 32
  • The drink-pushing task tests whether the policy combines short-distance pushes into moving a novel object from the table bottom to its top edge; Table 3 lists each task’s progress-score definition.Figure 22 gives examples of the expanded action-generalization benchmark.Source: §B.2.1.1 · p. 33, Table 3 · p. 33, Figure 22 · p. 33

B.2.1.2 Task generalization progress score definition … B.2.3.1 Task progress score for in-distribution, visual, semantic and action generalization tasks

This appendix defines task-generalization benchmarks and task-progress scores for ALOHA, Bi-arm Franka, and Apollo, covering in-distribution and varied generalization settings. It describes benchmark scope and identifies the tables that define scores for each robot and scene.

B.2.1.2 Task generalization progress score definition

  • The ALOHA task-generalization benchmark covers 12 tasks across multiple scenes, with instructions, objects, and initial conditions unseen in training.Figure 23 illustrates task execution, and Table 4 defines progress for each task.Source: §B.2.1.2 · p. 33, Figure 23 · p. 34, Table 4 · p. 35

B.2.2. Bi-arm Franka robot

  • The Bi-arm Franka benchmark comprises 44 tasks: 20 in-distribution and 24 variants across instruction, visual, and action generalization.It spans three scenes: a workbench for manipulation and tool interaction, a computer scene for dexterous cable and peripheral tasks, and a NIST board for belt assembly and removal; it includes cable insertion and NIST assembly tasks.Source: §B.2.2 · pp. 34–35, Figure 24 · p. 35

Workbench scene

  • Progress scores for Bi-arm Franka workbench tasks are defined in Tables 5 and 8, covering workbench and semantic generalization.Table 5 covers the Workbench scene; Table 8 covers semantic generalization on the workbench.Source: Table 5 · p. 37, Table 8 · p. 39

Computer scene

  • Progress scores for Bi-arm Franka computer tasks are defined in Tables 6 and 9, covering computer-scene and semantic generalization.Table 6 covers the Computer scene; Table 9 covers semantic generalization on the computer scene.Source: Table 6 · p. 38, Table 9 · p. 40

NIST Assembly Task Board 2

B.2.3. Apollo humanoid robot

  • The Apollo humanoid benchmark includes 24 tasks with instruction, visual, and action variants, using the ALOHA task-generalization tasks.Figure 26 illustrates variations for measuring instruction, visual, and semantic generalization.Source: §B.2.3 · p. 40, Figure 26 · p. 40

B.2.3.1 Task progress score for in-distribution, visual, semantic and action generalization tasks

  • Apollo task-progress scores are defined for in-distribution and visual generalization in Table 10, semantic generalization in Table 11, and action generalization in Table 12.The tables specify progress scores for the corresponding task categories.Source: Table 10 · p. 41, Table 11 · p. 41, Table 12 · p. 42

B.2.3.2 Qualitative Results … B.3.2. ALOHA benchmark →Bi-arm Franka and Apollo humanoid robot →Bi-arm Franka

Gemini Robotics 1.5 generalizes to unseen objects, receptacles, and surface heights on Apollo, while the cross-embodiment benchmarks test tasks on robots different from those used to collect the data. The benchmark passages describe task selection and progress-score definitions, but provide no score values.

B.2.3.2 Qualitative Results

  • On Apollo, Gemini Robotics 1.5 grasps novel objects, places them in unseen receptacles, and manipulates objects at unseen heights using automatically emerging grasp strategies.Humanoid control is especially challenging because of whole-body degrees of freedom, multi-finger dexterity, and limited workspace visibility requiring active perception.Source: §B.2.3.2 · p. 43, Figure 27 · p. 43

B.3. Cross-embodiment benchmark

  • The cross-embodiment benchmark tests tasks on one of three robots when data were collected only on another robot, and vice versa.Source: §B.3 · p. 44

B.3.1. Bi-arm Franka →ALOHA benchmark

  • Figure 28 illustrates execution of cross-embodiment tasks in the Bi-arm Franka →ALOHA benchmark.Source: Figure 28 · p. 44

B.3.2. ALOHA benchmark →Bi-arm Franka and Apollo humanoid robot →Bi-arm Franka

  • ALOHA task selection includes drawer opening and precise closure of a pear-shaped organizer, motions unavailable in Bi-arm data, plus easier packing tasks to test transfer and performance.Source: §B.3.2 · pp. 44–45

B.3.1. Bi-arm Franka →ALOHA benchmark

  • Table 13 presents progress scores for the Bi-arm Franka →ALOHA benchmark.Source: Table 13 · p. 45

B.3.2. ALOHA benchmark →Bi-arm Franka and Apollo humanoid robot →Bi-arm Franka

  • The Humanoid-to-Bi-arm Franka benchmark covers 11 tasks, with Figures 29 and 30 overviewing them and Table 14 defining their progress scores.Source: §B.3.2 · p. 45

B.3.3. ALOHA benchmark →Humanoid robot … B.5.1. Success rate for generalization performance

The appendix defines cross-embodiment and multi-step benchmark tasks and progress scoring, and reports Section 3 success rates; Figure 35 describes GR 1.5 as outperforming baselines across four variation types.

B.3.3. ALOHA benchmark →Humanoid robot

B.4. Multi-step benchmark

  • Multi-step benchmarks combine tasks into compound or abstract goal-oriented instructions, with compound tasks often requiring a particular order for a perfect score.The Apollo example requires packing gifts in a specified sequence; abstract instructions include “pack all the gifts.”Source: §B.4 · p. 48

B.4.1. ALOHA robot

B.4.2. Bi-arm Franka robot

B.4.3. Humanoid robot

B.5. Success rate

  • This section reports success rates for all results in Section 3.Source: §B.5 · p. 54

B.5.1. Success rate for generalization performance

  • Figure 35 describes GR 1.5 as consistently outperforming baselines and handling all four types of generalization variation more effectively.Source: Figure 35 · p. 54

B.5.2. Success rate for data and model ablation · B.5.3. Success rate for thinking ablation · C. Gemini Robotics-ER 1.5 is a generalist embodied reasoning model

The data-and-model ablation reports that Gemini Robotics 1.5 consistently outperforms baselines trained on single- or multi-robot data without the Motion Transfer recipe. The supplied context also describes Gemini Robotics-ER 1.5 as establishing a state of the art in embodied reasoning; the thinking ablation figure compares multi-step benchmark success rates with and without thinking, but provides no result values here.

B.5.2. Success rate for data and model ablation

  • Gemini Robotics 1.5 consistently outperforms baselines trained on single- or multi-robot data without the Motion Transfer recipe.Figure 36 summarizes the dataset and training-recipe ablation on the authors’ robots.Source: Figure 36 · p. 55

C. Gemini Robotics-ER 1.5 is a generalist embodied reasoning model

  • Gemini Robotics-ER 1.5 establishes a state of the art in embodied reasoning, including visual and spatial understanding, task planning, and progress estimation.These are the embodied-reasoning capabilities identified in the supplied paper context.Source: Figure 36 · p. 55

B.5.3. Success rate for thinking ablation

  • The multi-step benchmark thinking ablation compares success rates with and without thinking enabled during inference.The supplied figure caption does not state the comparative outcome or success-rate values.Source: Figure 37 · p. 56

C.1. Evaluation Details: Generality · C.2. Evaluation Details: Pointing · C.3. Additional Examples

The evaluation details describe how Gemini Robotics-ER 1.5 and other multimodal models were assessed across academic embodied-reasoning benchmarks, complex-pointing subtasks, and sampled reasoning examples.

C.1. Evaluation Details: Generality

  • Gemini Robotics-ER 1.5 and other multimodal models were evaluated on 15 academic embodied-reasoning benchmarks, with aggregate results in Fig. 8 and individual results in Tables 19 and 20.Gemini 2.5 Flash graded text-based VQA response accuracy for both multiple-choice and freeform questions.Source: §C.1 · p. 57
  • Gemini 2.5 and GPT-5 were accessed from September 1–20, 2025, with default thinking budgets and no tool use.Source: §C.1 · p. 57
  • Table 19 covers model performance across the 15 academic embodied-reasoning benchmarks; its GPT-5 and GPT-5-mini results came from September 2025 API access.The passage provides the table's scope but no benchmark cell values.Source: Table 19 · p. 57

C.2. Evaluation Details: Pointing

  • Table 21 breaks down the evaluation of complex pointing by subtask.The supplied passage identifies the breakdown but does not state subtask scores.Source: §C.2 · p. 57, Table 21 · p. 58

C.3. Additional Examples

  • Figure 38 presents sampled thoughts from GR-ER 1.5 while it performs embodied-reasoning tasks.Source: §C.3 · p. 57, Figure 38 · p. 59
  • Table 20 reports model performance on MMMU, GPQA, and Aider Polyglot, with GPT-5 and GPT-5-mini API results using default thinking settings and no tools.Those API results were obtained in September 2025; the passage does not provide cell values.Source: Table 20 · p. 58

Related papers

Loading 2510.03342v3…