Green-VLA: Staged Vision-Language-Action Model for Generalist Robots

• Deployment remains challenging because robot data vary in observations, action spaces, sampling rates, and quality, while behavior cloning saturates on long-horizon objectives and autoregressive reasoning can add prohibitive control latency.Source: §1 · p. 1
• Green-VLA filters and smooths demonstrations using DataQA quality measures, aligns speed temporally, balances sampling, and adds speed-conditioned control and joint-prediction guidance for precise targeting.Source: §1 · pp. 2–3
• Its five-stage recipe moves from multimodal grounding through multi-embodiment pretraining and embodiment adaptation to RL refinement, using a shared action layout to support transfer across robot types.The framework supports humanoids, mobile manipulators, and collaborative arms.Source: §1 · p. 2
• The same policy controls robots from single-arm manipulators to the 32-DoF Green humanoid without architectural changes, including coordinated bimanual and fine-grained fingertip control.Source: §1 · p. 2
• Green-VLA matches or exceeds prior pretrains at R0, remains competitive after embodiment tuning, and gains most after RL alignment, especially in long-horizon success, recovery, and precise task following.Source: §1 · p. 3
• The paper identifies multilingual instruction following, low-latency task-decomposition reasoning, and embodied memory with trajectory replay as future directions for practical long-horizon deployment.Source: §1 · p. 3

L0–L1

• L0 starts from a large-scale image/video–text VLM without robot actions; L1 adds multimodal web knowledge for semantic grounding, physics, affordances, and task structure.L0 lacks refined visual, physical, and multi-view understanding needed for robots.Source: §2 · p. 3

R0

• R0 uses multi-embodiment robotics data to learn affordance priors and feasible action distributions while abstracting embodiment-specific quirks into cross-domain invariants.In Green-VLA, R0 is the repository of base manipulation skills and is intended to improve data efficiency across embodiments.Source: §2 · p. 3

R1

• R1 adapts general pretraining to a target embodiment for immediate success-rate gains without new large-scale data, balancing tuning, architecture, and inference efficiency.Techniques include hyperparameter and optimizer search, embodiment-aware state/action heads, and faster inference designs.Source: §2 · p. 3

R2

• R2 addresses behavior cloning’s long-horizon and contact-rich credit-assignment and OOD-action limits with reward-based RL alignment, improving success rate, average chain length, and robustness.Green-VLA combines BC priors with task rewards, failure recovery, and preference-like feedback for stability and long-horizon robustness.Source: §2 · pp. 3–4

3.1 Data Pipeline

• Training combines 24M multimodal web samples for L1 with 184M robotics-domain samples spanning over 3,000 hours for R0, aligning RGB and proprioception by physical progress.The L1 data includes VQA, pointing, bounding-box prediction, pixel-wise trajectory prediction, multi-view QA, captioning, and spatial reasoning; sampling uses custom task weights.Source: §3.1 · p. 4, §3.2.1 · pp. 4–5

3.2.1 Multi-modal Web Data

• L1 draws on web datasets for spatial reasoning, pointing, robotics-related VQA, and multi-view question answering, with task-specific sampling weights.The web-data mixture includes RefSpatial, RoboPoint, PixMo-Points, and OpenSpaces, among other sources.Source: §3.2.1 · pp. 4–5

3.2.2 Robotics data

• R0 combines open-source and internally collected demonstrations across humanoids, mobile manipulators, and fixed-base arms, with varied camera views, proprioceptive states, tasks, and skills.The self-collected Green Humanoid dataset records 32 joint states and three fisheye cameras over 48 hours; an ALOHA dataset adds 11.2 hours of pick-and-place demonstrations.Source: §3.2.2 · pp. 5–6
• Mirroring and selectively reversing 48 hours of Green Humanoid recordings yielded 167 effective training hours, with reversible tasks and instructions transformed to match the new trajectories.Mirroring swaps and flips left/right camera streams and transforms corresponding joint trajectories; time reversal excludes inherently irreversible interactions.Source: §3.2.2 · p. 7

3.3 Data Quality Assurance

• DataQA filters episodes for missing cameras or frames, implausible duration, insufficient motion, erratic movement, blurry imagery, and task-inconsistent gripper patterns.Image sharpness is estimated from local Laplacian-boundary scores; pick-and-place episodes, for example, are checked for an expected open-closed-open gripper pattern.Source: §3.3 · p. 7
• Dataset sampling weights reflect size, visual diversity, and trajectory smoothness; AgiBot twofinger, Galaxea, and DROID receive the highest weights within their respective dataset groups.Galaxea’s base movement raises its state-diversity measure relative to other ALOHA datasets.Source: §3.3 · p. 8

4.1 Architecture

• A vision–language encoder fuses images, robot state, and instructions, while a flow-matching expert predicts embodiment-aware action chunks for a single low-latency multi-robot policy.The architecture uses efficient attention, lightweight heads, and fewer denoising steps for real-time deployment.Source: §4.1 · p. 9

4.2 Task planner module

• A GigaVision planner parses requests, decomposes physical goals into symbolic subtasks, and uses task-completion feedback to advance or replan; it remains frozen and runs only at inference.Non-action queries receive text or speech replies without invoking the low-level policy.Source: §4.2 · p. 9

4.3 Unified Action Space for Multi-Embodiment Control

• Green-VLA maps robot actions into a 64-dimensional fixed semantic layout and masks unused slots, preventing padding-related loss while enabling consistent cross-embodiment action semantics.Structured embodiment and control-type prompts specify the active action parameterization.Source: §4.3 · p. 10
• Retargeting maps demonstrations into meaningful target-humanoid actions, while a uniform-to-weighted dataset-sampling curriculum helps preserve early learning across embodiments.Retargeting matches corresponding limbs and grasp intent rather than naively padding or duplicating differing degrees of freedom.Source: §4.3 · p. 11
• Optical-flow-based resampling aligns motion speeds across datasets, and speed-conditioned modulation lets inference trade local precision for faster long-horizon progress without retraining.Training warps trajectories at sampled speed factors, while the inference speed parameter controls temporal resolution.Source: §4.3 · pp. 11–12
• The policy predicts episode progress for downstream subgoal updates and uses a Gaussian-mixture OOD detector [27] to correct trajectories toward training-state regions.When predicted state density falls below a threshold, actions are recomputed using a density-gradient correction.Source: §4.3 · p. 12

4.4 Guidance with joint prediction module

• For unfamiliar instructed objects, Green-VLA’s joint prediction head uses the current observation and language instruction to predict a workspace target; inverse kinematics finds a feasible joint configuration, and ΠGDM guides flow-matching actions toward it.Source: §4.4 · pp. 12–13

4.5 RL fine-tuning (R2)

• R2 uses Q-guided trajectory optimization with validation and source-distribution optimization by a separate actor, improving candidate actions while leaving base-model weights unchanged.Optimized trajectories are added to the R1 data and native fine-tuning is repeated; the noise-distribution actor supports controlled exploration near training actions.Source: §4.5 · pp. 13–14

5 Experiment Metrics Across Phases

• Green-VLA is trained on approximately 3,000 hours of unified demonstrations, versus over 10,000 hours for π0, enabling assessment of quality alignment and unified actions under constrained data.The latest configuration is approximately 5B parameters and uses Qwen3-VL-4B-Instruct with a dedicated flow-matching action expert. R0 training exceeds 105 optimization steps on 64 H100 GPUs.Source: §5 · pp. 14–15
• R0 Green-VLA achieves higher task-following accuracy and success rate than π0 on Cobot Magic, and multi-embodiment pretraining performs strongly on trained datasets without embodiment-specific tuning.On Simpler benchmarks, the R0 model outperforms other policies at the same pure-pretraining stage and performs comparably to several fine-tuned baselines.Source: §5 · p. 15

5.1 R1: Efficient Fine-Tuning

• In e-commerce shelf picking, JPM predicts a language-conditioned target point and guidance steers flow matching toward it, boosting success across coarse-category, exact-SKU, and out-of-distribution recognition regimes.The evaluation also measures top-1 correctness, task success, and completion time under randomized object poses and distractors; near-identical items and packaging changes make fine-grained recognition challenging.Source: §5.1 · pp. 16–17, Figure 11 · p. 17
• Green-robot evaluations show robust instruction following across pick, place, handover, fruit sorting, and table cleaning, including out-of-distribution scene layouts.The protocol randomizes objects, distractors, and clutter; it measures exact task success, average chain length for sequences up to four primitive actions, and table-cleaning time.Source: §5.1 · pp. 17–19, Figure 12 · p. 17

5.2 R2: RL Alignment

• On Simpler BRIDGE WidowX, R2 RL alignment raises R1 success rate by an absolute 24%, while CALVIN results show gains in long-horizon consistency, error recovery, compositional success, and average chain length.On CALVIN, Green-VLA R1 and CALVIN-fine-tuned π0 are comparable, with R1 slightly ahead on aggregate success rate and multi-step chains; R2 improves over both. R2 e-commerce fine-tuning also improves approach trajectories, contact geometry, and resistance to drops or slips on challenging items.Source: §5.2 · p. 19

6 Conclusion

• The framework combines demonstration quality alignment, embodiment-aware action unification, balanced multi-robot training, conservative RL, and inference-time guidance.DataQA filters and smooths heterogeneous demonstrations and uses optical-flow-based resampling for temporal alignment; embodiment prompts support positive transfer, while target-balanced sampling stabilizes multi-embodiment flow matching. Inference optimizations and guidance enable low-latency control, including for novel language-specified items.Source: §6 · p. 19
• Green-VLA outperforms prior foundation policies at comparable pretraining stages, approaches fine-tuned baselines on Simpler and CALVIN, and succeeds in bimanual and OOD-layout humanoid deployments.Source: §6 · p. 19
• With R2 RL alignment, Green-VLA achieves state-of-the-art results on Simpler BRIDGE WidowX and competitive, near-state-of-the-art performance on CALVIN ABC→D.Source: §6 · p. 20
• Performance remains dependent on retargeting fidelity, residual dataset bias, and dexterous-skill coverage; proposed extensions target multilingual following, real-time reasoning-control, and safety-aware online RL.Source: §6 · p. 20

• VLA contributors were I. Apanasevich, M. Artemyev, R. Babakyan, P. Fedotova, D. Grankin, E. Kupryashin, A. Misailidi, D. Nerus, A. Nutalapati, and G. Sidorov.Source: §7 · p. 20
• RL fine-tuning contributors were I. Efremov, M. Gerasyov, D. Pikurov, and Y. Senchenko; data pipeline contributors were S. Davidenko, D. Kulikov, and M. Sultankin.Source: §7 · p. 20
• Control contributors were K. Askarbek, O. Shamanin, D. Statovoy, E. Zalyaev, and I. Zorin; data collection contributors were A. Letkin, E. Rusakov, A. Silchenko, and V. Vorobyov.Source: §7 · p. 20
• S. Sobolnikov contributed to benchmarks, and A. Postnikov was the project supervisor.Authors are listed in alphabetical order.Source: §7 · p. 20