PaperlayerView exact revision on arXiv

Source-linked AI summary

AtomVLA: Scalable Post-Training for Robotic Manipulation via Predictive Latent World Models

Xiaoquan Sun, Zetian Xu, Chen Cao, Zonghe Liu, Yihan Sun, Jingrui Pang, Ruijian Zhang, Zhen Yang, Kang Pang, Dingxin He, Mingqi Yuan, Jiayu Chen

arXiv:2603.08519v1cs.RO

TL;DR

Current VLA approaches rely on coarse, high-level instructions, leaving limited explicit intermediate guidance for long-horizon tasks. AtomVLA decomposes demonstrations into atomic subtasks and uses a predictive world model for offline post-training, achieving 97.0% success on LIBERO and 48.0% on LIBERO-PRO.

  • Problem

    Current VLA approaches rely on high-level instructions, while explicit intermediate guidance for complex long-horizon tasks remains limited.

    Source: §I · p. 2

  • Method

    AtomVLA uses an LLM to decompose demonstrations into atomic subtasks and a predictive world model to score candidate action chunks against subtask goals for offline post-training.

    Source: §I · p. 2, B. Reward Modeling and RL Post-training · p. 3

  • Results

    97.0% success on LIBERO and 48.0% on LIBERO-PRO; real-world evaluation highlights performance in long-horizon manipulation of deformable objects.

    Source: §I · p. 2

  • Takeaways & Limitations

    Real-world evaluations on the Galaxea R1 Lite support AtomVLA’s applicability to tasks requiring prolonged execution sequences.

    Source: §VI · p. 8

Abstract

from arXiv · show

Vision-Language-Action (VLA) models demonstrate remarkable potential for generalizable robotic manipulation. The execution of complex multi-step behaviors in VLA models can be improved by robust instruction grounding, a critical component for effective control. However, current paradigms predominantly rely on coarse, high-level task instructions during supervised fine-tuning. This instruction grounding gap leaves models without explicit intermediate guidance, leading to severe compounding errors in long-horizon tasks. Therefore, bridging this instruction gap and providing scalable post-training for VLA models is urgent. To tackle this problem, we propose \method, the first subtask-aware VLA framework integrated with a scalable offline post-training pipeline. Our framework leverages a large language model to decompose high-level demonstrations into fine-grained atomic subtasks. This approach utilizes a pretrained predictive world model to score candidate action chunks against subtask goals in the latent space, mitigating error accumulation while significantly improving long-horizon robustness. Furthermore, this approach enables highly efficient Group Relative Policy Optimization without the prohibitive expenses associated with online rollouts on physical robots. Extensive simulations validate that our AtomVLA maintains strong robustness under perturbations. When evaluated against fundamental baseline models, it achieves an average success rate of 97.0\% on the LIBERO benchmark and 48.0\% on the LIBERO-PRO benchmark. Finally, experiments conducted in the real world using the Galaxea R1 Lite platform confirm its broad applicability across diverse tasks, especially long-horizon tasks. All datasets, checkpoints, and code will be released to the public domain following the acceptance of this work for future research.

I. INTRODUCTION · II. RELATED WORK · A. Vision-Language-Action Models

AtomVLA targets the instruction-grounding and policy-refinement gaps that limit robust long-horizon VLA deployment. It decomposes demonstrations into atomic subtasks and uses predictive latent-world scoring to support offline post-training.

I. INTRODUCTION

  • VLA deployment is constrained by computational and data demands, weak grounding of sparse language amid visual tokens, compounding long-horizon errors, and costly, risky online reinforcement learning.Source: §I · p. 2
  • Lightweight models [7], [8],, semantic-reasoning methods [10], and simulation data, [12] address parts of the problem, but imitation learning and reactive policies impede generalization and offline optimization.Source: §I · p. 2
  • AtomVLA is a two-stage post-training framework that uses an LLM to decompose high-level demonstrations into atomic subtasks, aligning semantic guidance with action chunks.Source: §I · p. 2
  • A V-JEPA2-based predictive latent world model scores candidate action trajectories against subtask goals and supplies reward signals for RL post-training without costly real-world interaction or pixel-level synthesis.Source: §I · p. 2
  • 97.0% success on LIBERO and 48.0% on LIBERO-PRO; real-world Galaxea R1 Lite experiments also demonstrate success on complex long-horizon deformable-object tasks.The cited passage highlights folding a T-shirt as an example.Source: §I · p. 2

A. Vision-Language-Action Models

B. Reinforcement Learning for Post-Training · III. ATOMVLA · A. Model Architecture

AtomVLA combines atomic subtask instructions with predictive latent world-model scoring and GRPO for scalable offline post-training. Its architecture pairs a multimodal backbone with a flow-matching action head, while LLM-generated annotations divide demonstrations into basic manipulation subtasks.

B. Reinforcement Learning for Post-Training

  • AtomVLA combines V-JEPA latent-world-model scoring with GRPO to refine policies offline, avoiding costly physical-robot interactions and pixel-level prediction errors in long sequences.The paper presents this pipeline as providing a reliable reward signal and stable improvements on long-horizon tasks.Source: B. Reinforcement Learning for… · p. 2
  • AtomVLA uses an LLM to decompose high-level demonstrations into atomic subtask instructions, then uses a world model for RL post-training.The figure contrasts this approach with typical VLA models that rely on SFT training.Source: Figure 2 · p. 3

A. Model Architecture

  • The VLM backbone is Qwen3-VL-4B-Instruct [29], which encodes multiview observations and high-level and subtask instructions into features conditioning the action expert via cross-attention.The backbone is coupled with an action head.Source: A. Model Architecture · p. 3
  • The cross-attention Diffusion Transformer action head predicts horizon-N action chunks as a flow-matching task, regressing the target velocity of Gaussian-noised ground-truth actions.Training samples Gaussian noise and a continuous time value, mixes the noise with the ground-truth chunk, and minimizes error between predicted and target velocity.Source: A. Model Architecture · p. 3
  • At inference, the action head initializes a chunk from Gaussian noise and refines it through K explicit Euler updates to produce the final action chunk.Each update uses step size 1/K and the action expert conditioned on the visual and instruction features.Source: A. Model Architecture · p. 3
  • For each demonstration, an LLM uses sampled video frames and task context to generate subtask instructions with start and end frames, constrained to 2–5 basic manipulation actions.GPT-4o performs the decomposition and annotation; the prompt targets actions such as picking up, placing, opening or closing, and pushing objects.Source: A. Model Architecture · p. 3

B. Reward Modeling and RL Post-training

The section presents an offline post-training approach that scores candidate action chunks against intermediate and final goals using a pretrained latent world model, avoiding costly online rollouts. Group-normalized rewards and KL regularization support policy updates anchored to expert demonstrations.

Reward design

  • Candidate action chunks are scored against subtask-boundary and final-episode goals in latent space, while an imitation-deviation term discourages implausible actions and reward hacking.Subtask-boundary frames serve as intermediate goals; the final frame supplies the episode goal. The subgoal and final-goal terms provide stage-wise guidance and long-horizon consistency, respectively.Source: B. Reward Modeling and RL Post-training · p. 3

Action-conditioned world model

  • A predictor rolls out candidate action chunks from current-observation latents encoded by frozen V-JEPA2 [13], enabling predicted consequences to be compared with goal states.The action-conditioned latent dynamics model is built on a pretrained video representation.Source: B. Reward Modeling and RL Post-training · p. 3

GRPO optimization

C. Training Pipeline · IV. SIMULATION EXPERIMENTS

The pipeline combines demonstration-based supervised fine-tuning with offline, world-model-guided GRPO post-training. The simulation section frames questions about benchmark performance, post-training gains, robustness factors, and real-world deployment.

C. Training Pipeline

  • Offline GRPO samples demonstration states, generates candidate action chunks, and scores subgoal and final-goal rewards with the V-JEPA2 world model.The policy is initialized from the SFT checkpoint.Source: C. Training Pipeline · p. 4
  • Post-training applies GRPO updates with a KL constraint to the frozen SFT policy, mainly updating the action expert to preserve stable vision-language representations and improve long-horizon robustness.Source: C. Training Pipeline · p. 4
  • SFT jointly optimizes the Qwen3-VL backbone and action expert using flow-matching loss to supervise predicted action chunks from robot demonstrations.Source: C. Training Pipeline · p. 4
  • To reduce gradient variance, training repeats each batch with independently sampled (τ, ϵ) and uses mixed precision.Source: C. Training Pipeline · p. 4

IV. SIMULATION EXPERIMENTS

  • Simulation experiments ask whether AtomVLA outperforms existing VLA models on standard and long-horizon benchmarks and whether world-model-guided GRPO consistently improves performance.Source: §IV · p. 4
  • The evaluation agenda also examines how subtask instruction refinement and action chunk size affect long-horizon robustness, and whether AtomVLA can be deployed effectively on real-world robots.Source: §IV · p. 4

A. Experimental Setup · B. Results Analysis

Experiments on LIBERO and perturbation-focused LIBERO-PRO report gains from post-training and atomic subtask instructions, while in-domain training and action chunking also support strong performance. The results emphasize robustness under perturbations and improved long-horizon instruction grounding.

A. Experimental Setup

  • Evaluation used success rate over 50 trials per task on LIBERO and LIBERO-PRO, comparing AtomVLA with representative VLA baselines.LIBERO covers spatial, object, goal, and long-horizon suites; LIBERO-PRO adds controlled perturbations in objects, scenes, and instructions. SFT trained on LIBERO for two epochs; post-training sampled 100K trajectories with 10 candidates per sample.Source: A. Experimental Setup · pp. 4–5

B. Results Analysis

  • +6.0% on LIBERO-Goal was AtomVLA’s largest gain over SFT, while post-training improved performance across all LIBERO suites.GRPO post-training uses group-level comparative supervision and world-model evaluation to reinforce candidate rollouts that better satisfy intermediate subgoals and the final goal.Source: B. Results Analysis · p. 5
  • 93% average success on LIBERO was achieved without cross-embodiment large-scale pretraining, using in-domain multitask demonstrations.The passage reports that cross-embodiment pretraining incurs higher training costs and yields inconsistent downstream gains due to action-space discrepancies and uneven data quality.Source: B. Results Analysis · p. 5
  • AtomVLA retained stable, non-trivial success across LIBERO-PRO perturbations, including position and task-related shifts where many baselines fell near zero.The authors attribute this robustness to SFT’s imitation prior combined with post-training that reinforces goal-consistent action selection under world-model evaluation.Source: B. Results Analysis · p. 5
  • -9.6% on LIBERO-Long followed language-instruction removal; task instructions reached 90.0% success, and adding subtask descriptions raised it to 92.2%.Subtask instructions provide an intermediate objective while preserving the overall goal, helping focus actions and improve credit assignment over long horizons.Source: Table I · p. 5, B. Results Analysis · p. 5
  • Chunk size 4 yielded optimal manipulation performance; overly large chunks reduced execution flexibility and exacerbated compounding errors.Action chunking is used to improve temporal coherence and implicit planning in generated actions.Source: Table II · p. 5, B. Results Analysis · p. 5

C. Ablation Study · V. REAL-WORLD EXPERIMENTS

The ablation favors combining intermediate subtask guidance with final-goal consistency, while the real-world task visualization spans basic manipulation and longer-horizon folding tasks.

C. Ablation Study

  • Using either elementary reward separately improves average LIBERO success by approximately 3.0% over SFT.Source: C. Ablation Study · p. 5
  • The combined reward achieves a 4.0% overall improvement over SFT on LIBERO.The study reports the combined reward as the most stable configuration.Source: C. Ablation Study · p. 5
  • On LIBERO’s Long suite, the combined reward improves success by 4.4% over SFT.Source: C. Ablation Study · p. 5
  • The combined reward mitigates noise from long-horizon predictions while avoiding short-sighted biases from focusing only on immediate subgoals.Source: C. Ablation Study · p. 5

V. REAL-WORLD EXPERIMENTS

  • The real-world task visualization depicts stacking bowls, placing fruit in a basket, hanging a cup, and opening a drawer.Source: Figure 4 · p. 6
  • The visualization also depicts hard, long-horizon real-world tasks of folding a T-shirt and a towel.Source: Figure 4 · p. 6

A. Experimental Setup

The real-world experiments evaluate AtomVLA on a stationary Galaxea R1 Lite platform across basic and hard manipulation tasks, under standard and perturbed generalization settings.

Experimental platform

Task categories

Evaluation settings

  • Evaluation uses Standard conditions matching training demonstrations and Generalization conditions that introduce four deployment perturbations.These perturbations are randomized object and receptacle positions, unseen distractors, varied target heights, and semantically equivalent instruction variants.Source: A. Experimental Setup · p. 7

Training and evaluation

  • Each task has 100 demonstrations; training uses 2 SFT epochs followed by post-training on 100K trajectories with 10 candidates per state, and evaluation runs 20 trials per task.For Generalization trials, conditions are randomly distributed across the four perturbation types.Source: A. Experimental Setup · p. 7

B. Results Analysis · VI. CONCLUSION

AtomVLA improves robustness under generalization variations and combines LLM-driven subtask decomposition with predictive-world-model-guided offline post-training. The conclusion reports gains on simulation benchmarks and real-world long-horizon tasks, while identifying static subtask boundaries as a limitation.

B. Results Analysis

  • 47.5% average success rate versus π0 [6] at 29.2% under GE variations, an 18.3% absolute improvement.For Fold T-shirt and Fold towel, AtomVLA achieves 25% and 35% success rates, respectively, in GE Hard tasks involving deformable objects.Source: B. Results Analysis · p. 7, Figure 6 · p. 7
  • Subtask instruction decomposition and world-model-guided post-training mitigate error accumulation in highly unstructured environments.The reported robustness advantage is especially pronounced in Hard deformable-object tasks requiring precise long-horizon control.Source: B. Results Analysis · p. 7

VI. CONCLUSION

  • AtomVLA addresses the instruction-grounding gap and inefficient offline reinforcement learning in long-horizon robotic tasks.LLM-driven subtask decomposition provides sequential guidance for complex operations.Source: §VI · p. 8
  • The two-stage training paradigm combines supervised fine-tuning with scalable offline GRPO guided by a predictive world model, achieving gains on LIBERO and LIBERO-PRO.The conclusion attributes the framework’s performance gains to this combination of training stages.Source: §VI · p. 8
  • Real-world Galaxea R1 Lite evaluations highlight AtomVLA’s advantages on tasks requiring prolonged execution sequences, including deformable-object manipulation such as folding a T-shirt.The Results Analysis reports robustness under GE variations, with 25% success on Fold T-shirt and 35% on Fold towel, far exceeding the baseline.Source: B. Results Analysis · p. 7, §VI · p. 8
  • Adapting to highly dynamic environments remains challenging because the current system relies on static LLM-generated subtask boundaries.The limitation is stated for the current framework despite its ability to enable long-horizon tasks.Source: §VI · p. 8
  • Future work will explore end-to-end subtask generation during execution and world models that perceive uncertainty.The paper also plans to improve generalization to entirely unseen objects and novel semantic instructions.Source: §VI · p. 8

APPENDIX · A. Example Figure · B. Training Details

The appendix supplements the main paper with task-segmentation examples and training configurations for LIBERO and Galaxea R1 Lite experiments.

APPENDIX

  • The appendix outlines four parts: LIBERO segmentation examples, training and post-training configurations, subtask-generation prompts and formats, and baseline-model descriptions.It says the training details include GPU configurations, learning rates, and input specifications.Source: APPENDIX · p. 10

A. Example Figure

  • Figure 7 presents atomic-subtask examples from the LIBERO dataset.Source: Figure 7 · p. 10

B. Training Details

  • Table VII lists SFT-training hyperparameters for the LIBERO dataset.Source: Table VII · p. 11
  • Table VIII lists post-training hyperparameters for the LIBERO dataset.Source: Table VIII · p. 11
  • Table IX lists SFT-training hyperparameters for the real-world dataset on the Galaxea R1 Lite platform.Source: Table IX · p. 11
  • Table X lists post-training hyperparameters for the real-world dataset on the Galaxea R1 Lite platform.Source: Table X · p. 11

C. Prompt for fine-grained atomic task generation

The prompt guides a vision-language model to segment sampled tabletop-robot videos into temporally complete, coarse-grained atomic subtasks aligned with the overall goal. It standardizes subtask granularity, wording, and frame-indexed JSON output.

C. Prompt for fine-grained atomic task generation

  • The model must sequence the episode’s main steps from sampled frames and frame indices, covering the timeline without gaps or overlaps and interpreting actions in light of the global task goal.The prompt frames this as precise temporal segmentation of tabletop robotic manipulation.Source: C. Prompt for fine-grained atomic task… · p. 12
  • Instructions use standardized forms such as “Place [object] into/on [target]” and “Move arm to [location],” with specific object names grounded in the global task context.Examples include distinguishing objects by attributes or location, such as “white mug” or “left plate.”Source: C. Prompt for fine-grained atomic task… · p. 12
  • The output is strict JSON with an instruction and start/end frames for each segment, whose boundaries are estimated using the sampled images’ supplied frame indices.The prompt provides examples of consecutive frame ranges for pickup and placement segments.Source: C. Prompt for fine-grained atomic task… · p. 12

D. Baseline.

The comparison evaluates eleven VLA baselines spanning generalist, spatially aware, reasoning-based, and flow-matching approaches to robotic manipulation.

D. Baseline.

  • OpenVLA combines a Llama 2 backbone with DINOv2 and SigLIP, while TraceVLA adds visual trace prompting and Octo-base supports language and goal-image conditioning.OpenVLA was pretrained on 970k trajectories, TraceVLA was fine-tuned on 150K trajectories, and Octo-base was trained on approximately 800K trajectories.Source: D. Baseline · p. 13
  • SpatialVLA injects explicit 3D information for manipulation, while CoT-VLA predicts future visual frames as intermediate goals before generating action sequences.SpatialVLA uses Ego3D Position Encoding and Adaptive Action Grids; CoT-VLA is a 7B model built for temporal planning.Source: D. Baseline · p. 13
  • π0 uses flow matching over a pretrained VLM, whereas ThinkAct separates high-level reasoning from low-level control through a visual latent plan.ThinkAct reinforces embodied reasoning plans with action-aligned visual rewards before using the latent plan to condition its action policy.Source: D. Baseline · p. 13
  • NORA targets efficient real-time manipulation, while NORA-1.5 adds a flow-matching action expert and reward-driven post-training with action-conditioned world models.NORA uses a Qwen-2.5-VL-3B backbone and FAST+ tokenizer; NORA-1.5 uses deviation-from-ground-truth heuristics for direct preference optimization.Source: D. Baseline · p. 13
  • MolmoAct integrates depth-aware perception, spatial planning, and action prediction, while X-VLA uses soft prompts to capture cross-embodiment variation.The 7B-D MolmoAct variant is described as performing across simulation and real-world tasks; X-VLA-0.9B is described as achieving state-of-the-art performance across simulations and real-world robots.Source: D. Baseline · p. 13

Related papers

Loading 2603.08519v1…