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
TL;DR
VLA models need better intermediate instruction grounding and scalable post-training for robust long-horizon manipulation. AtomVLA decomposes demonstrations into atomic subtasks and uses a predictive latent world model for offline policy optimization, achieving strong benchmark and real-world results.
Problem
Existing VLA training relies on coarse high-level instructions and faces compounding errors in long-horizon tasks, while online post-training is costly to scale.
Method
AtomVLA uses LLM-generated atomic subtasks and a predictive latent world model to score candidate action chunks and provide offline reinforcement-learning signals.
Results
AtomVLA achieves 97.0% success on LIBERO and 48.0% on LIBERO-PRO, with real-world Galaxea R1 Lite evaluations highlighting long-horizon manipulation.
Takeaways & Limitations
Subtask guidance and latent-world-model post-training support robust long-horizon manipulation without relying on costly online physical-robot rollouts.
Abstract
from arXiv · showhide
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
AtomVLA addresses limitations in scaling and long-horizon robustness by combining fine-grained subtask decomposition with predictive-world-model-guided post-training. It reports strong benchmark performance and real-world applicability to complex manipulation.
- Current VLA methods face prohibitive pretraining costs and difficulties transferring lab demonstrations to robust long-horizon deployment.
- AtomVLA is a scalable two-stage post-training framework for robust VLA models in long-horizon robotic tasks.
- An LLM decomposes high-level demonstrations into atomic subtasks, providing explicit stage-wise guidance aligned with action chunking.
- A predictive latent world model evaluates candidate action trajectories against subtask goals, producing reward signals without costly physical rollouts or pixel-level synthesis.
- 97.0% success on LIBERO and 48.0% on LIBERO-PRO demonstrate strong benchmark performance, while Galaxea R1 Lite experiments include complex long-horizon deformable-object manipulation.
B. Reinforcement Learning for Post-Training
AtomVLA combines VLA action prediction with subtask-conditioned inputs and offline reinforcement-learning post-training. Its design uses language-generated temporal subtasks and action chunks refined through model-based latent guidance.
- World-model-based post-training is motivated by the prohibitive cost and limited scalability of online reinforcement-learning rollouts on robots.
- Figure 2 contrasts typical SFT-based VLA training with AtomVLA’s language-model decomposition and world-model RL post-training.
- The VLM backbone encodes multiview observations, high-level instructions, and subtask instructions into features that condition the action expert.
- The diffusion-transformer action head predicts horizon-N action chunks using flow matching with Gaussian-noised demonstrations.
- An LLM decomposes demonstrations into subtasks with natural-language instructions and start/end frames, using standardized basic manipulation actions and 2–5 subtasks per task.
B. Reward Modeling and RL Post-training
AtomVLA uses a pretrained latent world model to score candidate action chunks against subtask and final goals, then applies offline GRPO to refine the policy without online robot rollouts.
- Motivation: SFT alone can underperform on challenging long-horizon tasks, motivating a scalable offline signal anchored to expert demonstrations.The method treats subtask boundary frames as intermediate goals for reward construction.
- World-model reward: A frozen V-JEPA2 encoder and action-conditioned predictor roll out candidate futures in latent space for goal-based scoring.The predictor estimates latent future representations from the current observation and candidate action chunk.
- World-model reward: The reward combines subtask and final-goal energies with imitation deviation, discouraging implausible actions and reward hacking.Lower goal energies are better; the final-goal term preserves long-horizon consistency.
- GRPO post-training: GRPO normalizes rewards across candidate groups into advantages and increases the likelihood of higher-advantage action chunks.A KL penalty to a frozen SFT reference policy stabilizes updates.
- GRPO post-training: Post-training updates only the action-head parameters rather than the full model.
C. Training Pipeline
AtomVLA trains in two stages: supervised fine-tuning with high-level and atomic subtask instructions, followed by offline GRPO post-training guided by the reward model.
- SFT Training: SFT jointly trains the Qwen3-VL backbone and action expert on demonstrations conditioned on observations, task instructions, and subtask instructions.The predicted action chunk is supervised with a flow-matching loss.
- Post Training: Offline GRPO starts from the SFT checkpoint and uses the world-model reward to refine action generation.
- Post Training: The pipeline targets stable vision-language representations while improving long-horizon robustness.
- Evaluation questions: The experiments evaluate benchmark performance, world-model-guided post-training, instruction refinement and chunk size, and real-world deployment.These questions correspond to Q1 through Q4.
A. Experimental Setup
The evaluation combines LIBERO and perturbed LIBERO-PRO simulations with ablations of post-training, instruction inputs, and action-chunk horizons, plus real-world assessment.
- Benchmarks: LIBERO includes four suites covering spatial, object, goal, and long-horizon generalization, while LIBERO-PRO adds controlled perturbations.Each LIBERO suite contains 500 expert demonstrations across 10 tasks.
- Evaluation protocol: The post-training setup samples 100K trajectories and evaluates 10 candidates per sample.
- Post-training evaluation: Post-training improves AtomVLA over SFT across all LIBERO suites, with the largest reported gain on LIBERO-Goal at +6.0%.The comparison attributes this pattern to group-level comparative supervision and world-model evaluation.
- Instruction ablation: Removing language instruction reduces LIBERO-Long performance by 9.6%, while task instructions reach 90.0% and subtask augmentation reaches 92.2%.Subtask descriptions supply an intermediate objective while preserving the overall task goal.
- Action-chunk ablation: Action chunking improves planning and sequence stability, with chunk size 4 producing the best manipulation performance.Overly large chunks reduce execution flexibility and worsen compounding errors.
C. Ablation Study
The ablation study finds that combining subgoal and final-goal rewards provides the most stable LIBERO performance, while the tables frame benchmark comparisons and the figure illustrates evaluated real-world tasks.
- Approximately 3.0%: each elementary reward improves average LIBERO success over the SFT baseline.
- 4.0%: the combined reward improves overall LIBERO performance over SFT.
- 4.4%: the combined reward improves performance on the more challenging LIBERO-Long suite over SFT.
- The combined reward balances intermediate subtask guidance with final-goal consistency, reducing long-horizon prediction noise and short-sighted subgoal bias.
- Figure 4 contrasts basic manipulation tasks with hard, long-horizon folding tasks involving a T-shirt and towel.
A. Experimental Setup
The experiments evaluate AtomVLA on a stationary Galaxea R1 Lite platform across basic and hard real-world manipulation tasks, with standard and generalization settings introducing controlled variations.
- Platform and tasks: The Galaxea R1 Lite setup uses a dual-arm mobile platform with two 6-DoF arms, wrist cameras, and a head camera.
- Platform and tasks: Basic tasks include stacking bowls, placing fruit in a basket, hanging a cup, and opening a drawer.
- Platform and tasks: Hard tasks involve folding a T-shirt or folding a towel twice to reach target configurations.
- Evaluation settings: The evaluation uses Standard conditions matching training demonstrations and Generalization conditions designed to assess robustness.
- Evaluation settings: Generalization varies object positions, introduces unseen distractors, changes target heights, and paraphrases task instructions.
B. Results Analysis
Real-world results show comparable standard-setting capabilities but a clear generalization advantage for AtomVLA, especially on long-horizon deformable-object tasks.
- Generalization results: AtomVLA maintains greater robustness than π0 under spatial, visual, and instruction variations.
- Generalization results: 47.5%: AtomVLA’s average Generalization success rate exceeds π0’s 29.2% by 18.3 percentage points.
- Hard tasks: 25% and 35%: AtomVLA succeeds on Generalization Fold T-shirt and Fold towel tasks, respectively.
- Overall findings: The conclusion reports gains on LIBERO and LIBERO-PRO alongside real-world reliability for prolonged manipulation sequences.
APPENDIX
The appendix supplies visual task-segmentation examples and training or post-training configurations for LIBERO and the Galaxea R1 Lite platform.
- Supplementary materials: The supplementary material includes visual examples of temporal task segmentation into coarse-grained atomic subtasks.
- Supplementary materials: Figure 7 visualizes atomic subtasks on the LIBERO dataset.
- Training configurations: The appendix provides SFT and post-training hyperparameters for the LIBERO dataset.
- Training configurations: It also provides SFT and post-training hyperparameters for the real-world Galaxea R1 Lite dataset.
C. Prompt for fine-grained atomic task generation
The prompt generates contiguous, coarse-grained atomic subtasks from sampled robot-video frames and the overall task goal, using standardized language and temporal boundaries.
- Prompt for Instruction Generation: The model segments tabletop manipulation videos into a contiguous sequence of coarse-grained atomic subtasks aligned with the global task context.It uses sampled video frames, frame indices, and the overall goal to identify the episode’s main steps.
- Prompt for Instruction Generation: The granularity rules merge Approach plus Grasp into Pick up and Move plus Release into Place.
- Prompt for Instruction Generation: Typical pick-and-place episodes are represented with 2-4 main steps.
- Prompt for Instruction Generation: Generated instructions use object-specific templates such as placing an object into or on a target and moving an arm to a location.The prompt requires specific object names matching the global task context.
- Prompt for Instruction Generation: The output is a JSON object containing each subtask’s instruction, start frame, and end frame.The prompt also requires sampled-image indices to estimate temporal boundaries accurately.
D. Baseline.
The comparative evaluation uses OpenVLA and TraceVLA as baseline VLA models with distinct architectural and representation choices.
- Baseline: OpenVLA is a 7B-parameter VLA combining a Llama 2 language backbone with DINOv2 and SigLIP visual representations.It is pretrained on 970k real-world robot trajectories from the Open X-Embodiment dataset.
- Baseline: TraceVLA is a spatial-temporal enhanced VLA that uses visual trace prompting to encode state-action trajectories.