GigaBrain-0.7: Scaling Embodied Foundation Models to Emergent Capabilities with a Three-System Architecture

GigaBrain Team, Angen Ye, Axiang Sun, Can Jin, Chenxi Cheng, Chong Shi, Dengke Shang, Dingqian Zhang, Guan Huang, Guangqiang Wang, Guangqing Ding, Guo Li, Hangcong Li, Hengyu Zhong, Hongtao Lu, Jianbo Qin, Jiming Mao, Jing Zhu, Jindi Lv, Jingzhi Cui, Junjie Xie, Junyi Bao, Kai Liu, Lei Yuan, Limin Long, Lv Feng, Mingming Yu, Peng Li, Pengfei Yi, Qi Li, Qianli Zhang, Qingfang Li, Qitang Hu, Rui Zhang, Shaoyan Sun, Shibo Sun, Shiying Duan, Tenghui Chen, Tianze Liu, Weijie Ke, Wenyao Xue, Xiaofeng Wang, Xiaoyu Tian, Xinyu Liu, Xinze Chen, Yang Wang, Yankai Wang, Yejun Zeng, Yifan Li, Yifei Nie, Yilong Li, Yilong Liu, Yongchao Feng, Yumeng Wang, Yun Ye, Zhichao Liu, Ziheng He, Zonghai Yang, Zheng Zhu

GigaBrain-0.7 unifies vision-language understanding, predictive world modeling, and continuous robot control in a three-system architecture.

How can a three-system architecture—separating visual-language planning, predictive evaluation, and action control—enable embodied agents to scale to complex, long-horizon tasks?

Embodied agents struggle to scale across heterogeneous robot platforms because reactive policies often fail to anticipate future states or coordinate long-horizon tasks. GigaBrain-0.7 introduces a three-system architecture that separates high-level planning, predictive world modeling, and continuous action generation, all trained in a single-stage joint optimization. The model scales to over 37,000 hours of multi-embodiment data, achieving significant gains in zero-shot instruction following and task success rates over prior state-of-the-art models.

Paper Primer

The core mechanism hinges on a three-system design: System 2 (Understanding and Planning) decomposes instructions into subtasks; System 3 (Prediction and Evaluation) uses a world model to generate future subgoal images and progress values; and System 1 (Action and Control) fuses these signals to generate continuous action chunks. The systems communicate through structured semantic and value-based interfaces, allowing the model to ground actions in both current visual context and anticipated future states.

GigaBrain-0.7 demonstrates superior zero-shot generalization and task completion compared to preceding GigaBrain series and $\pi\$0.5.

Systematic scaling studies and real-world evaluations on Maker H01 and AgileX PiPER platforms. Substantial improvements in foundation capabilities and post-training success rates across diverse home and industrial scenarios.

To prevent the degradation of general vision-language capabilities during embodied training, the authors employ Soft Knowledge Insulation (Soft KI). This technique attenuates action gradients entering the vision-language backbone, allowing the model to adapt to physical control while preserving its semantic reasoning abilities.

Why is a three-system architecture necessary instead of a single monolithic policy?

A single policy often struggles to coordinate long-horizon tasks and reactive control simultaneously. By separating planning, prediction, and action, the model can explicitly reason about task progress and future visual states, which provides better conditioning for continuous motor control.

How does the model handle the diversity of 16 different robot morphologies?

The model uses a unified robot representation where proprioceptive states and Robot IDs are projected into a shared space. While the Action Expert parameters are shared, per-embodiment input/output projections handle specific kinematic differences and degrees of freedom.

GigaBrain-0.7 shifts the paradigm from reactive observation-to-action prediction toward a coordinated learning system that integrates predictive world modeling and experience-driven improvement.

Introduction and Motivation

We frame the need for predictive, multi‑system embodied intelligence beyond reactive VLA models.

Vision‑Language‑Action (VLA) models have become the dominant paradigm for general‑purpose robot control, yet they remain constrained by three intertwined gaps: (1) scaling to vastly larger, more heterogeneous data regimes, (2) integrating diverse robot embodiments without destructive interference, and (3) lacking mechanisms to anticipate future states before committing to actions.

The core problem is to turn reactive VLA pipelines into a unified learning system that can ingest heterogeneous trajectories, reason about future outcomes, and continuously improve across many robot morphologies.

**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.

The key shift is moving from simple reactive control toward a predictive, coordinated embodiment where understanding, planning, and evaluation are learned together.

Related Work

We situate GigaBrain‑0.7 among prior VLA architectures, training regimes, and world‑model integrations.

Vision‑Language‑Action (VLA) research clusters into three architectural families. Autoregressive VLM‑as‑Actor models such as RT‑2, OpenVLA, FAST, Galaxea G0.5, and OpenVLA‑OFT generate continuous actions as token streams, but decoding cost grows with horizon and dimensionality. Serial cascaded designs (GR00T N1, Gemini Robotics, Qwen‑RobotManip) keep a dedicated Action Expert separate from the VLM, while Parallel Mixture‑of‑Transformers ($\pi$₀, $\pi$₀.5, $\pi$₀.7, GigaBrain‑0, Wall‑OSS‑0.5, HyVLA‑0.5, Xiaomi‑Robotics‑0/1) interleave a continuous Action Expert more tightly with the vision‑language backbone.

Training pipelines also diverge. Early VLA systems adopt a multi‑stage recipe: a pretrained VLM is first adapted to token‑level action prediction (RT‑2, OpenVLA) and later augmented with a continuous flow‑matching expert ($\pi$₀.5, Xiaomi‑Robotics‑0/1). Recent work unifies vision‑language and action objectives in a single stage (GigaBrain‑0, Wall‑OSS‑0.5, LingBot‑VLA 2.0), jointly optimizing multimodal cross‑entropy, discrete action, and continuous flow matching. Post‑training refines these joint models for specific robots and tasks, and reinforcement‑learning extensions (VLA‑RL, SimpleVLA‑RL, Z‑1, RL Token) further adapt policies from simulated or real rollouts.

World‑model integration provides complementary benefits. Predictive video models generate synthetic trajectories (DreamGen) or serve as unified world‑action generators (GR‑2, DreamZero, Unified World Models, GigaWorld‑Policy, Cosmos Policy/3). Closed‑loop loops such as World‑VLA‑Loop let policies improve the world model and vice‑versa. Predictive embeddings (JEPA‑VLA) and value‑derived progress signals (Feat2Go, RAMP in GigaBrain‑0.5M*) condition policy learning, a pattern GigaBrain‑0.7 extends with its dedicated System 3 for future‑state prediction and evaluation.

A VLA model ingests visual observations and language instructions, then directly emits robot‑level control signals.

**Table 1.** Composition of the embodied trajectory corpus after cleaning.

Data Curation and Scaling

We describe how we assemble and clean a multi‑source corpus for training GigaBrain‑0.7.

Training VLA models on action data alone erodes their vision‑language abilities; we therefore curate complementary vision‑language data and a rigorous cleaning pipeline.

**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.

We turn heterogeneous robot and human demonstrations into a single, clean representation so the model can learn from all sources without being confused by format or quality mismatches.

Convert the raw CSV log into LeRobot v3.0 JSON, creating fields for episode ID, 20 frames, joint angles, and camera image paths.

Arrange the joint vector as [left‑arm‑j1, left‑arm‑j2, right‑arm‑j1, right‑arm‑j2, head‑pan, head‑tilt, base‑x, base‑y] and pad missing entries with zeros and a mask flag.

Rewrite the original instruction “pick up the red block” with the LLM to “Grasp the red block on the table” and add subtask tags “approach → grasp → lift”.

Extreme‑value filter removes frame 3 where joint velocity exceeds the 99th percentile (a sensor glitch).

Stationary removal merges frames 7–9 (no joint change) into a single idle segment.

Cleaning eliminates outlier spikes and long idle periods, yielding a compact, high‑signal trajectory that the model can learn from efficiently.

How does this pipeline differ from simply concatenating all raw datasets?

Simple concatenation would mix incompatible file formats, mismatched joint orderings, and noisy or duplicated frames. Our pipeline first normalizes every sample to a common schema, aligns robot kinematics, rewrites instructions into a unified language, and then prunes outliers and idle periods, guaranteeing that the training data are homogeneous and informative.

**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.

The Three-System Architecture

GigaBrain‑0.7’s three‑system design couples prediction, planning, and control into a single trainable pipeline.

Long‑horizon robot interaction demands more than a single observation‑action pair: the policy must retain temporal context, anticipate future states, and judge whether its current behavior is making progress.

The agent is split into three cooperating subsystems: System 2 plans from vision‑language input, System 3 predicts what will happen and how valuable that future is, and System 1 turns all signals into concrete robot actions.

System 3 predicts a future observation and a scalar value that measures task progress, turning raw visual dynamics into a usable signal for the controller.

The core trick is to let the pretrained vision‑language backbone attend to its own tokens while the Action Expert attends jointly to both streams, preserving the VLM’s autoregressive interface yet enriching it with action‑relevant context.

Step 1: Causal self‑attention on $[v_1, v_2]$ produces updated vectors $v'_1, v'_2$ (each attends only to earlier tokens).

Step 2: Concatenate $[v'_1, v'_2, a_1]$ and run bidirectional attention with $a_1$ as queries; the resulting context vector $c$ mixes information from both $v'$s and $a_1$.

Step 3: Apply the Action Expert feed‑forward network to $c$, yielding the new action‑expert hidden state $a'_1$.

Step 4: The updated pair $(v'_1, v'_2, a'_1)$ is passed to the next layer.

This toy walk‑through shows how the Action Expert can read the full semantic context while the vision‑language stream remains causally constrained, which is the essence of the joint‑attention trick.

Joint attention update for one transformer layer

Temporal‑Spatial Blocks give System 1 a short‑term visual memory without blowing up the token budget, and Soft Knowledge Insulation protects the pretrained vision‑language backbone from destabilizing gradients originating in the continuous‑action pathway.

Experimental Setup

Key scaling results show dual‑stream architecture and System 3 signals dramatically boost robot success.

Dual‑Stream coupling yields 30 % shirt‑folding success, while Last‑Layer and Multi‑Layer variants achieve 0 %.

Table 5 reports 30 % for Dual Stream versus 0 % for the other two architectures.

The experiments also reveal that scaling pretraining data consistently lowers validation loss and improves real‑robot success, especially for complex deformable‑object tasks.

**Figure a.** Validation loss under different pretraining-data scales.

**Figure b.** Real-robot performance as robot-data scale increases.

**Figure c.** Effect of UMI and EGO data in the pretraining mixture.

**Figure a.** Without temporal context. The policy repeatedly revisits similar interaction states and falls into a repeated-action cycle.

System 3’s future‑state and value conditioning further improve both success rates and efficiency, as shown by the ablation study.

Across all three tasks (Clothes folding, Gift wrapping, Cube sorting) the full System 3 configuration (+SubImage + Value) consistently yields the highest scores and fastest completions.

The table presents performance metrics for three different backbones: PaliGemma2 (3.5B), Qwen3.5 (5B), and Gemma 4 (8.5B). The tasks evaluated are "Clean Desk", "Fruit Picking", and "Shirt Folding".

**Table 5.** Tab. 5 shows a clear trade-off between action-generation capability and computational efficiency. Restricting the action expert to the final VLM representation yields the lowest inference latency, whereas the multi-layer cross-attention variant incurs additional computational cost without improving task success in the evaluated settings. In contrast, the dual-stream architecture achieves the strongest real-robot performance across all three tasks, including nonzero success on the more challenging shirt-folding task, although at a higher inference cost.

System 3 Analysis

We probe System 3’s components and post‑training gains via ablations and out‑of‑the‑box tests.

Recall that System 3 predicts future state values, letting the agent evaluate actions before committing.

Removing both image conditioning (+SUBIMAGE) and value conditioning (+VALUE) from System 3 drops the average task score by 12 points.

Figure 14 shows the base variant scores 73 pts while the combined subImage + value variant scores 85 pts.

**(b) With temporal context.** Historical observations help disambiguate the revisited state, allowing the policy to exit the repeated interaction and continue 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 b.** Maker H01: grapes are absent from the H01 training data, and neither the evaluation scene nor the task composition was observed during 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 a.** AgileX PiPER: folding unseen garment instances from diverse unstructured initial states.

**Figure b.** Maker H01: folding an unseen garment instance from an unstructured initial state.

**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.

Benchmark Evaluations

Benchmark results show GigaBrain‑0.7 leads across vision‑language and complex manipulation tasks.

GigaBrain‑0.7 achieves the highest average success rate (84.9 %) on the post‑training complex manipulation benchmark.

Table 7 reports an overall average of 84.9 % for GigaBrain‑0.7, exceeding all other evaluated models.

Table 6 evaluates six instruction‑conditioned manipulation tasks on AgileX PiPER and Maker H01, covering color, target, and direction discrimination. GigaBrain‑0.7 attains competitive success rates (75–85 %) across all six tasks, matching or surpassing earlier baselines on most columns.

Table 7 expands the evaluation to seven complex manipulation tasks. While GigaBrain‑0.7’s per‑task scores are modest on some structured tasks, its average of 84.9 % remains the top across both robot embodiments, highlighting consistent gains on contact‑rich and multi‑stage interactions.

Table 8 reports embodied vision‑language performance on the MiMo‑Embodied suite. GigaBrain‑0.7 outperforms all compared models on both the Spatial Understanding (average 78 %) and Affordance (average 81 %) groups, yielding the highest overall mean across the 13 benchmarks.

**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 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.

**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.

Simulation Benchmark Results

GigaBrain‑0.7 leads across three simulation benchmarks, setting new performance highs.

GigaBrain‑0.7 outperforms all publicly available VLA models on three complementary simulation benchmarks, achieving the highest scores on each metric.

Table 9 shows the top Hard‑setting success (67.35 %) on RoboTwin 2.0; Table 10 reports the highest overall success rate (46.1 %) and score (42) on EBench; Table 11 records the best spatial‑reasoning score (0.68) on RoboColiseum.

EBench is a simulation suite that tests mobile manipulation, long‑horizon execution, and precise interaction, reporting both a task‑success rate and a progress‑score.

**Table.** Performance comparison of various models across Spatial Avg., Affordance Avg., and Overall metrics.

**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 $\pi_0$, $\pi_{0.5}$, 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.

This table compares the performance of various models based on their Success Rate (SR) and Score.

Experience-Driven Reinforcement

Experience‑driven reinforcement adds successive offline and online learning stages to a pretrained policy.

GigaBrain‑0.7 organizes embodied intelligence into three interacting systems, with System 3 predicting the value of future states before actions are taken.

The agent first learns a task‑specific policy from demonstrations, then repeatedly improves it by harvesting its own rollouts (offline) and by receiving human corrections while executing the refined policy (online).

Mean progress $\bar r = (0.2+0.5+0.8)/3 = 0.5$.

Compute raw advantages: $a_1 = 0.2-0.5 = -0.3$, $a_2 = 0$, $a_3 = 0.3$.

Exponentiate with $\beta=2$: $w_1 = e^{-0.6}\approx0.55$, $w_2 = e^{0}=1$, $w_3 = e^{0.6}\approx1.82$.

Normalize weights to sum to 1: $w_1'=0.20$, $w_2'=0.36$, $w_3'=0.44$.

The policy gradient is then a weighted sum of the three segment gradients, giving the highest influence to the most progressive rollout.

Advantage weighting turns raw progress into a smooth priority signal, so even modestly successful rollouts still contribute useful gradients.

How does the offline stage differ from standard offline RL that simply imitates past actions?

Standard offline RL treats every logged action as equally trustworthy. Here, each segment is re‑scored by a learned progress estimator, and the loss is weighted by the advantage of that progress. Thus the algorithm prefers trajectories that move the robot forward, while still learning from failures that contain informative partial progress.

**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.

Experience‑driven reinforcement lifts average task success from 30 % after SFT to 100 % after online RL.

Table 13 shows every task reaches 100 % success once the online stage is applied.

**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).

Appendix

Appendix provides full leaderboard snapshots and URLs for reproducibility of simulation benchmarks.

This appendix records the exact leaderboard states used for the simulation comparisons in Section 6.8.2, ensuring that future readers can verify the reported numbers.

RoboTwin 2.0 leaderboard was captured from https://robotwin-platform.github.io/leaderboard on August 15, 2026.

EBench leaderboard was captured from https://internrobotics.shlab.org.cn/eval/ on August 15, 2026; it includes submissions such as Qwen‑RobotManip that are omitted from the main‑text tables.

RoboColiseum leaderboard was captured from https://robocoliseum.ai/ on August 15, 2026; four capability dimensions are reported separately in Figures 30–33.

**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.

**Overall Ranking — clean2random (hard)**

**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.

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers