Vision-language-action (VLA) models have become a dominant paradigm for generalist embodied agents, demonstrating strong complex and long-horizon task completion in structured settings. Yet it remains an open question whether current VLA systems can benefit from more effective architectural design, scale to substantially larger and more heterogeneous data regimes, and achieve broader generalization across tasks and embodiments. To this end, we present GigaBrain-0.7, an embodied foundation model with substantially improved generalization across diverse robot embodiments. Specifically, GigaBrain-0.7 unifies understanding, prediction, and action through a three-system architecture, scales pretraining to over 37,000 hours of heterogeneous embodied data, and introduces one-stage alignment training that jointly optimizes vision-language understanding and multi-embodiment action generation. Compared with the preceding GigaBrain-0 series and prior state-of-the-art models including π0.5, GigaBrain-0.7 achieves substantial improvements in foundation zero-shot capabilities, language-conditioned instruction following, and post-training task success rates. In particular, on our in-house Maker H01 platform and mainstream robot embodiments, GigaBrain-0.7 demonstrates strong task adaptability and completion ability across both home and industrial scenarios. All training code and pretrained model weights will be released.
Figures & tables
Figure 1: Overview of GigaBrain-0.7 . GigaBrain-0.7 is an embodied foundation-model system that coordinates understanding and planning, prediction and evaluation, and action and control through a three-system architecture. It learns from heterogeneous pretraining data and on-robot experience to support generalization and continual improvement across diverse robot embodiments and real-world scenarios.
Data Type
Duration (h)
Proportion
Real Robot
20,535.65
55.12%
UMI
8,251.83
22.15%
EGO
2,862.36
7.68%
Simulation
1,453.92
3.90%
WM
4,153.22
11.15%
Total
37,256.98
100.00%
Table 1: Composition of the embodied trajectory corpus after cleaning.
Figure 2: Composition of the GigaBrain-0.7 training corpus. The embodied trajectory corpus contains 37.3K hours of real-robot, UMI, EGO, simulation, and world-model-generated data, while the VLM corpus contains 271,976,674 image–text/question-answering samples. The figure further summarizes representative robot embodiments, data-capture devices, atomic actions, and task/environment distributions.
Robot Type
Episodes
Frames
Frame Prop.
Agibot-G1
226,675
380,282,116
18.30%
AgileX
133,101
179,727,199
8.65%
Maker H01
866,864
910,496,160
43.82%
Ark
30,559
35,960,476
1.73%
Galaxea R1 Lite
27,793
35,058,414
1.69%
Franka
87,148
25,149,233
1.21%
Table 2: Statistics of the real-robot data by robot type.
Data Type
Dataset
Duration (h)
UMI
In-house Maker U01
6,876.25
UMI
Jianzhi 10K
1,375.58
UMI Total
–
8,251.83
EGO
EgoDex
724.84
EGO
EgoVerse
835.77
EGO
WiYH
165.38
Table 3: Composition of the UMI and EGO datasets after cleaning.
Figure 3: Real-robot data processing pipeline. Heterogeneous robot data are converted into the LeRobot v3.0 format, mapped to a unified robot representation, standardized through language-instruction rewriting and subtask annotation, and filtered through multi-stage quality control before training.
Figure 4: UMI data processing pipeline. Multi-camera observations, 6-DoF handle poses, and gripper states are temporally aligned, filtered for trajectory and modality quality, and retargeted to the Maker H01 end-effector frame before simulation and sampled real-robot replay validation.
Figure 5: EGO data processing pipeline. First-person manipulation videos are filtered for interaction and video quality, supplemented with hand-pose and hand–object information when necessary, and standardized across EgoDex, EgoVerse, WiYH, and in-house data sources.
Figure 6: Architecture of GigaBrain-0.7. GigaBrain-0.7 organizes embodied intelligence into three interacting systems. System 2 performs visual-language understanding and task planning. System 3 introduces explicit prediction and evaluation by predicting future observations and estimating task-related values. System 1 integrates the current observation and short-term memory, subtask instructions, robot state, and predictive signals to generate executable actions through parallel discrete and continuous action pathways. The architecture further exposes interfaces for offline and online experience reinforcement, connecting perception, prediction, action, and subsequent policy improvement.
Figure 7: Experience Reinforcement Pipeline. The experience reinforcement pipeline consists of three stages: supervised fine-tuning, offline reinforcement, and online reinforcement. Supervised fine-tuning establishes basic competence on complex, fine-grained out-of-domain tasks. Offline reinforcement develops progress awareness and error recovery, while online reinforcement further improves fine-grained manipulation and task success.
Backbone
Model Size
Clean Desk
Fruit Picking
Shirt Folding
PaliGemma2
3.5B
50%
88%
30%
Qwen3.5
5B
60%
12%
0%
Gemma 4
8.5B
100% *
92%
0%
Table 4: Comparison of VLM backbones for System 1. The evaluated configurations differ in model size and image resolution; the comparison is used to determine the backbone for subsequent experiments rather than as a controlled model-scaling study. Model size denotes the total parameter count of the full model, including both the VLM backbone and the action expert. * Gemma 4 Clean Desk was evaluated under a limited-range position-generalization setting.
Architecture
Clean Desk
Fruit Picking
Shirt Folding
Train s/step
Infer. s
Dual Stream
50%
88%
30%
4.93
0.221
Last-Layer Cross Attention
20%
40%
0%
4.65
0.073
Multi-Layer Cross Attention
20%
36%
0%
6.59
0.108
Table 5: Comparison of VLM–action expert coupling architectures. All variants use PaliGemma2 as the VLM backbone. Success rates are measured on real robots; training and inference costs are reported for the corresponding configurations.
Figure 8: Scaling GigaBrain-0.7 with heterogeneous embodied data. (a) Increasing the amount of pretraining data consistently reduces validation loss under a fixed model configuration. (b) Increasing robot-data scale improves real-robot performance, with a stronger effect on the more challenging deformable-object task. (c) UMI and EGO provide complementary supervision to robot trajectories, and their combination yields the strongest overall performance.
Figure 9: Qualitative analysis of temporal context in repeated-state manipulation. Each rollout is represented by ten temporally ordered frames. Without temporal context, the policy repeatedly returns to visually similar interaction states and produces a recurring action pattern. With temporal context, recent interaction history provides additional information about the current stage of execution, enabling the policy to break the repeated-action cycle and continue toward the task goal.
Figure 10: Prediction and evaluation with the World Value Model. The model predicts both future visual observations and future state values. The inferred advantage decreases sharply when the imagined trajectory leads to failure and increases when the trajectory recovers toward successful task execution.
Figure 11: Ablation of the prediction and evaluation interfaces in System 3. The base VLA is compared with image conditioning ( +SubImage ), value conditioning ( +Value ), and their combination. Completion time is reported only when successful task completion is observed.
Figure 12: Out-of-the-box multi-task language following of GigaBrain-0.7. We evaluate the same pretrained policy on AgileX PiPER and Maker H01 under both in-distribution and out-of-distribution settings.
Figure 13: Object, spatial, and compositional generalization of the pretrained GigaBrain-0.7 policy. The same pretrained GigaBrain-0.7 policy is deployed on AgileX PiPER and Maker H01 without task-specific adaptation. Panel (a) evaluates a pepper instance that is absent from the corresponding training data; moreover, the two target containers were not collected together in the same task configuration and their relative placement is randomized at evaluation time. Panel (b) introduces a broader compound distributional shift on Maker H01: grapes are absent from the H01 training data, and neither the evaluation scene nor the corresponding task composition was observed during H01 data collection.
Figure 14: Out-of-the-box complex manipulation with GigaBrain-0.7. The evaluation covers longer-horizon manipulation under ID and OOD settings on AgileX PiPER and Maker H01.
Figure 15: Out-of-the-box garment-instance and deformable-state generalization. The pretrained GigaBrain-0.7 policy is deployed on garment-folding tasks on AgileX PiPER and Maker H01 without task-specific adaptation. The evaluated garment instances are absent from the corresponding training data and are initialized in diverse, highly unstructured configurations. Panel (a) shows two representative AgileX PiPER rollouts with different unseen garments and initial states, while Panel (b) shows a Maker H01 rollout under the same evaluation principle. Each row contains five temporally ordered stages from a single rollout, illustrating how the policy maintains sustained interaction as the deformable garment geometry evolves throughout execution.
Figure 16: Post-training multi-task language following. We compare GigaBrain-0.7 with representative VLA baselines on six manipulation tasks that probe color, target, and directional grounding across two robot embodiments.
AgileX PiPER
Model
Color
Target
Direction
Avg.
Button Push
Spoon Grasping
Bowl Pick-Place
Tableware Pick-Place
Fork Pick-Place
Wood Block Pick-Place
π0.5
100.0
100.0
90.0
87.5
75.0
80.0
88.8
GigaBrain-0.1
66.7
77.8
85.0
72.2
80.0
75.0
76.1
G0.5
88.9
94.4
50.0
100.0
85.0
70.0
81.4
Xiaomi-Robotics-1
66.7
83.3
50.0
98.6
75.0
60.0
72.3
Table 6: Post-training multi-task language-following evaluation. We evaluate six instruction-conditioned manipulation tasks on AgileX PiPER and Maker H01, covering three complementary language-following capabilities: color discrimination (button push and spoon grasping), target discrimination (bowl and tableware pick-and-place), and direction discrimination (fork and wood-block pick-and-place). Each cell reports success rate (%). Avg. denotes the mean across all six tasks and is reported only for models with complete evaluations. Best available results in each column are shown in bold .
Figure 17: Post-training language following on AgileX PiPER: color grounding. Given the instruction “Grab the orange spoon and put it on the plate.” , GigaBrain-0.7 identifies the language-specified object by color and executes the corresponding pick-and-place behavior. The ten frames show temporally ordered stages of a representative rollout.
Figure 18: Post-training language following on AgileX PiPER: spatial grounding. Given the instruction “Grab the block on the left side of the plate and put it in the bowl.” , the policy resolves the relative spatial reference, selects the corresponding block, and executes the instructed transfer. The ten frames show temporally ordered stages of a representative rollout.
Figure 19: Post-training language following on Maker H01: color grounding. In a scene containing multiple colored alternatives, GigaBrain-0.7 identifies the language-specified blue target and executes the corresponding manipulation behavior. The rollout illustrates sensitivity to fine-grained visual attributes on the humanoid embodiment.
Figure 20: Post-training language following on Maker H01: target grounding. Given the instruction “Pick up the fork and move it to the basket.” , the policy identifies the requested utensil and maintains the language-conditioned target throughout the subsequent manipulation sequence.
Figure 21: Post-training language following on Maker H01: spatial grounding. Given the instruction “Grab the fork on the left side of the plate and put it in the basket.” , the policy grounds the relative spatial relation before executing the instructed pick-and-place sequence.
Figure 22: Post-training complex manipulation. The evaluation covers manipulation tasks requiring sustained interaction, multi-stage sequencing, deformable-object handling, and tool use on AgileX PiPER and Maker H01.
AgileX PiPER
Model
Cube Sorting
Play-Dough Kneading
Tableware Arranging
Household Item Storage
Rice Sweeping
Avg.
π0.5
95.0
95.0
77.8
95.0
20.0
76.6
GigaBrain-0.1
90.0
95.0
38.9
75.0
25.0
64.8
G0.5
45.0
20.0
38.9
85.0
0.0
37.8
Xiaomi-Robotics-1
35.0
85.0
44.4
75.0
5.0
48.9
GigaBrain-0.7
95.0
100.0
94.4
95.0
40.0
84.9
Table 7: Post-training single-task complex manipulation evaluation. We evaluate post-trained policies on a diverse set of complex manipulation tasks across AgileX PiPER and Maker H01. The benchmark covers object sorting, deformable-object manipulation, tableware arrangement, household-item storage, sweeping, food preparation and heating, and dish cleanup. Each cell reports task success rate (%). Avg. denotes the mean over all evaluated tasks on the corresponding embodiment and is reported only for models with complete evaluations. Best available results in each column are shown in bold .
Figure 23: Post-training complex manipulation on AgileX PiPER: play-dough manipulation. The policy performs repeated contact-rich interactions with deformable material, requiring subsequent actions to adapt as the object geometry changes throughout execution. The ten frames show temporally ordered stages of a representative rollout.
Figure 24: Post-training complex manipulation on AgileX PiPER: tabletop sweeping. GigaBrain-0.7 uses a brush to manipulate distributed material across the tabletop, requiring sustained tool–object contact and coordinated motion over multiple stages of execution.
Figure 25: Post-training complex manipulation on Maker H01: household-item organization. The humanoid policy performs a sequence of object selection, grasping, transport, and placement behaviors to organize multiple items across the workspace. The rollout illustrates sustained multi-stage execution on a substantially different robot embodiment.
Figure 26: Post-training complex manipulation on Maker H01: block sorting. GigaBrain-0.7 performs repeated object selection and placement across successive manipulation stages, requiring the policy to maintain task progress as the scene changes after each interaction.
Model
Spatial Avg.
Affordance Avg.
Overall
Xiaomi-Robotics-0 [ 9 ]
.1372
.0687
.1108
Spirit-v1.5 [ 70 ]
.3524
.2995
.3320
Wall-OSS-0.5 [ 86 ]
.2273
.0367
.1540
G0.5-base [ 23 ]
.3890
.3956
.3916
Hy-Embodied-0.5-VLA-UMI [ 89 ]
.2532
.0000
.1558
GigaBrain-0.7
.5215
.3669
.4621
Table 8: Summary of embodied vision-language evaluation on MiMo-Embodied [ 34 ] . Spatial Avg. is averaged over eight Spatial Understanding benchmarks, and Affordance Avg. over five Affordance benchmarks. Overall denotes the mean over all 13 evaluated benchmarks. For GigaBrain-0.7, the reported comparison is obtained before MiMo data are incorporated into subsequent continued pretraining. Bold indicates the best result among the compared models.
Model
Method
Easy
Hard
Overall
π0 [ 5 ]
Single
46.42
16.34
31.38
Xiaomi-Robotics-0 [ 9 ]
Co-Train
62.9
18.2
40.55
X-VLA [ 90 ]
Co-Train
68.0
20.9
44.45
X-WAM [ 32 ]
Co-Train
70.0
25.8
47.90
π0.5 [ 62 ]
Co-Train
70.7
46.0
58.35
GigaBrain-0.7
Co-Train
66.8
67.9
67.35
Table 9: Evaluation on RoboTwin 2.0 [ 15 ] . Under the official Co-Train protocol, one policy is jointly post-trained on all 50 tasks using 50 clean demonstrations per task. Single denotes task-specific fine-tuning with a separate checkpoint for each task. Success rates (%) are reported under the Easy (clean) and Hard (domain-randomized) settings. Overall denotes the mean of Easy and Hard. Best results in each column are shown in bold .
Model
SR (%)
Score
π0 [ 5 ]
23.59
37
X-VLA [ 90 ]
23.72
35
InternVLA-A1 * [ 8 ]
23.90
36
π0.5 [ 62 ]
28.08
42
GigaBrain-0.7
33.30
46.1
Table 10: Evaluation on EBench [ 24 ] among publicly available VLA models. SR denotes the overall task success rate, and Score measures task progress. Results for π0 , π0.5 , X-VLA, and GigaBrain-0.7 follow the corresponding leaderboard evaluations. * The InternVLA-A1 result is reproduced from the EBench evaluation reported in Qwen-RobotManip [ 87 ] , rather than from the current EBench leaderboard. Best results among the compared models are shown in bold . The complete official leaderboard, including additional submissions not shown here, is provided in Appendix A.2 .
Model
Instruction Following
Spatial Reasoning
Robustness
General Manipulation
π0 [ 5 ]
.3680
.1300
.3130
.3470
Xiaomi-Robotics-0 [ 9 ]
.6460
.2230
.5460
.3070
GR00T N1.7 [ 51 ]
.6460
.2490
.5380
.4380
π0.5 [ 62 ]
.7460
.3560
.6130
.5820
ACoT-VLA [ 92 ]
.7570
.3970
.6220
.4770
GigaBrain-0.7
.8166
.4729
.6800
.6092
Table 11: Evaluation on RoboColiseum [ 1 ] . The benchmark evaluates four complementary dimensions of robot capability. We report representative open-source models with complete evaluations. Best results in each column are shown in bold .
Figure 27: Representative real-robot tasks for experience-driven reinforcement learning. (A) Gift Box Packing (PiPER-X). (B) Bearing Installation (H01). (C) Link Installation (PiPER). (D) Cable Tie Insertion and Tying (PiPER-X).
Task
SFT Baseline
Offline RL
Online RL
Link Installation (PiPER)
20%
40%
100%
Gift Box Packing (PiPER-X)
80%
90%
100%
Cable Tie Insertion and Tying (PiPER-X)
0%
40%
100%
Bearing Installation (H01)
20%
60%
100%
Table 12: Task success rates at different training stages.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 28: Snapshot of the official RoboTwin 2.0 leaderboard. The leaderboard reports the evaluation protocol together with performance under the clean Easy and domain-randomized Hard settings. GigaBrain-0.7 is evaluated under the official Co-Train protocol used for the comparison in Tab. 9 .
Figure 29: Snapshot of the complete official EBench leaderboard. The leaderboard reports overall task success rate and aggregate task score for all listed submissions at the time of manuscript preparation.
Figure 30: Snapshot of the RoboColiseum Instruction Following leaderboard. The leaderboard evaluates the ability of robot policies to follow language-conditioned instructions at the time of manuscript preparation.
Figure 31: Snapshot of the RoboColiseum Spatial Reasoning leaderboard. The leaderboard evaluates spatial reasoning capability at the time of manuscript preparation.
Figure 32: Snapshot of the RoboColiseum Robustness leaderboard. The leaderboard evaluates policy robustness under the corresponding RoboColiseum protocol at the time of manuscript preparation.
Figure 33: Snapshot of the RoboColiseum General Manipulation leaderboard. The leaderboard evaluates general manipulation capability at the time of manuscript preparation.
We present RynnBrain 1.1, a family of embodied foundation models spanning 2B, 9B, and 122B-A10B scales. Trained with a unified spatio-temporal and physically grounded framework, RynnBrain 1.1 supports embodied perception, spatial reasoning, localization, and planning. Compared with RynnBrain 1.0, it further introduces contact-point prediction across the model family and native 3D grounding for the 2B and 9B models, yielding representations and outputs that are more directly aligned with robot manipulation. We also develop RynnBrain-VLA with a unified cross-embodiment action space and embodiment-specific masking, and deploy it on Unitree G1, Astribot-S1, and Tianji-Wuji. RynnBrain 1.1 achieves strong results on embodied cognition, localization, and 3D grounding, with the 122B-A10B model outperforming all evaluated proprietary and open-source models on VSI-Bench, MMSI, and RefSpatial-Bench. Real-robot experiments show that RynnBrain-initialized policies outperform Qwen-based and representative generalist VLAs, while joint multi-task and multi-embodiment training improves process scores and success rates over per-task training.
Embodied intelligence is often studied through specialized models for individual tasks such as manipulation or navigation, resulting in fragmented capabilities and limited generalization across tasks, environments, and robot embodiments. In this work, we study whether heterogeneous embodied decision-making problems can be unified within a single vision-language-action model. We present Qwen-VLA, a unified embodied foundation model that extends Qwen's vision-language modeling stack from perception, understanding, and reasoning to continuous action and trajectory generation through a DiT-based action decoder. Qwen-VLA is trained with a large-scale joint pretraining recipe over diverse data sources, including robotics manipulation trajectories, human egocentric demonstrations, synthetic simulation data, vision-and-language navigation data, trajectory-centric supervision, and auxiliary vision-language data. To support multiple robot platforms, we introduce embodiment-aware prompt conditioning, where robot-specific textual descriptions specify the current embodiment and control convention. We further cast manipulation, navigation, and trajectory prediction into a unified action-and-trajectory prediction framework, enabling transferable visual grounding, spatial reasoning, and continuous action generation across robot morphologies, task families, and environments. Experiments on manipulation, navigation, and trajectory-centric benchmarks show consistent multi-task performance and out-of-distribution generalization under variations in scene layout, background, lighting, object configuration, and robot embodiment. Qwen-VLA-Instruct achieves 97.9% on LIBERO, 73.7% on Simpler-WidowX, 86.1%/87.2% on RoboTwin-Easy/Hard, 69.0% OSR on R2R, 59.6% SR on RxR, 76.9% average OOD success in real-world ALOHA experiments, and 26.6% zero-shot success on DOMINO dynamic manipulation.
Cross-embodiment transfer in vision-language-action (VLA) models remains challenging because low-level state and action spaces differ fundamentally across robot platforms. We observe that the high-level cognitive process underlying manipulation, including scene perception, object identification, task planning, and sub-task decomposition, is largely shared across embodiments. Based on this observation, we present ZR-0, a 2.6 billion parameter end-to-end VLA model that uses dense Embodied Chain-of-Thought (ECoT) supervision to align cross-embodiment representations within the vision-language model (VLM). ZR-0 adopts a dual-stream architecture: a pre-trained VLM (System 2) generates structured ECoT reasoning during training, while a Diffusion Transformer-based action expert (System 1) produces continuous action chunks via flow matching. The two components are coupled through cross-attention, with an attention mask that restricts the action expert to input prompt features only, enabling ECoT generation to be entirely skipped at inference without any performance loss. ZR-0 is pre-trained on ProcCorpus-60M, a large-scale dataset comprising approximately 60 million frames (approximately 1,000 hours) from over 400K trajectories, with dense ECoT annotations covering 96.8% of all frames. We evaluate ZR-0 on three simulation benchmarks spanning single-arm (LIBERO), bimanual (RoboTwin 2.0), and humanoid (RoboCasa GR-1 Tabletop) embodiments, as well as real-world experiments on the xArm platform, demonstrating strong performance across all settings. Code and model checkpoints are available at https://github.com/RUCKBReasoning/ZR-0.