AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots

Likui Zhang; Tao Tang; Zhihao Zhan; Xiuwei Chen; Zisheng Chen; Jianhua Han; Jiangtong Zhu; Pei Xu

cs.ROMar 8, 2026

AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots

Authors: Likui Zhang, Tao Tang, Zhihao Zhan, Xiuwei Chen, Zisheng Chen, Jianhua Han, Jiangtong Zhu, Pei Xu, +4 more

Organizations: Sun Yat-sen University · Yinwang Intelligent Technology Co. Ltd. · Peng Cheng Laboratory

Abstract

Recent advances in Visual-Language-Action (VLA) models have shown promising potential for robotic manipulation tasks. However, real-world robotic tasks often involve long-horizon, multi-step problem-solving and require generalization for continual skill acquisition, extending beyond single actions or skills. These challenges present significant barriers for existing VLA models, which use monolithic action decoders trained on aggregated data, resulting in poor scalability. To address these challenges, we propose AtomicVLA, a unified planning-and-execution framework that jointly generates task-level plans, atomic skill abstractions, and fine-grained actions. AtomicVLA constructs a scalable atomic skill library through a Skill-Guided Mixture-of-Experts (SG-MoE), where each expert specializes in mastering generic yet precise atomic skills. Furthermore, we introduce a flexible routing encoder that automatically assigns dedicated atomic experts to new skills, enabling continual learning. We validate our approach through extensive experiments. In simulation, AtomicVLA outperforms $π_{0}$ by 2.4% on LIBERO, 10% on LIBERO-LONG, and outperforms $π_{0}$ and $π_{0.5}$ by 0.22 and 0.25 in average task length on CALVIN. Additionally, our AtomicVLA consistently surpasses baselines by 18.3% and 21% in real-world long-horizon tasks and continual learning. These results highlight the effectiveness of atomic skill abstraction and dynamic expert composition for long-horizon and lifelong robotic tasks. The project page is here.

Figures & tables

Figure 1: Overview of AtomicVLA. Unlike previous VLA models with a single action head, which suffer from limited scalability and severe interference among mixed skills, AtomicVLA employs a SG-MoE architecture to build a scalable skill expert library. By unifying task planning and action execution within this framework, it achieves strong performance on long-horizon and continual learning tasks in both simulation and real-world settings.

Figure 2: (a) AtomicVLA Pipline. AtomicVLA is a framework that unifies task planning and action execution. The VLM adaptively predicts atomic skill abstraction and latent action. Action Decoder in the SG-MoE architecture receives both the latent action and the newly inferred atomic skill abstraction, and generates fine grained motor actions. (b) Skill-Guided Mixture of Experts. SG-MoE includes a skill router, a shared expert, and multiple atomic-skill experts. The router selects the top skill expert based on the atomic skill, and the action token is processed by both the activated skill expert and the shared expert. (c) Continual Learning with Skill Expansion. New skills are added by training only the new expert and extending the router. (d) Task Planning Embodied Data Generation. High-quality embodied reasoning data are generated using principal-axis analysis with InternVideo2.5 [ 52 ] model.

Figure 3: Inference Example of AtomicVLA. We visualize two tasks from LIBERO-LONG. For each task, the top row shows the task progression, and the bottom row shows AtomicVLA’s inferred outputs. Gray blocks denote Thinking, while colored blocks indicate Acting, with colors corresponding to the activated skill experts. The left row shows the initial task state (top) and the skill-expert activation during inference (bottom).

Method	Spatial	Object	Goal	Long	Avg.
Octo [ 47 ]	78.9	85.7	84.6	51.1	75.1
OpenVLA [ 22 ]	84.9	88.4	79.2	53.7	76.5
SpatialVLA [ 37 ]	88.2	89.9	78.6	55.5	78.1
CoT-VLA [ 65 ]	87.5	91.6	87.6	69.0	83.9
$\pi_{0}$ [ 3 ]	96.4	98.8	95.8	85.2	94.2
$\pi_{0.5}$ [ 17 ]	98.8	98.2	98.0	92.4	96.9

Table 1: Comparison of Different Methods on LIBERO Benchmark(%).

Method	Task	Tasks Completed in a Row					Avg. Len $\uparrow$
Method	Task	1	2	3	4	5	Avg. Len $\uparrow$
$\pi_{0}$ [ 3 ]	ABC $\to$ D	94.3	87.0	77.9	68.5	59.4	3.87
$\pi_{0.5}$ [ 17 ]	ABC $\to$ D	91.9	84.6	79.4	75.5	71.0	4.02
AtomicVLA (Ours)	ABC $\to$ D	95.0	87.8	81.9	75.0	69.1	4.09
AtomicVLA* (Ours)	ABC $\to$ D	94.1	88.7	85.2	81.7	77.6	4.27

Table 2: Long-horizon Robotic Manipulation Evaluation on CALVIN Benchmark(%).

Figure 4: Error Recovery Capability Demonstration. When encountering a skill execution failure, AtomicVLA automatically assesses the progress and re-executes the current skill.

Figure 5: Demonstrations show the execution process of AtomicVLA* (second row) and baselines $\pi_{0.5}$ (first row).

Method	InP	IntoD	IntoM	Avg.	$\Delta$ Avg.
$\pi_{0}$ [ 3 ]	45	55	10	36.7	–
$\pi_{0.5}$ [ 17 ]	65	35	35	45	–
AtomicVLA	65	60	45	56.7	$+20.0\uparrow$
AtomicVLA*	75	60	55	63.3	${+18.3}\uparrow$

Table 3: Long-horizon Multi-task Experiments(%). InP, IntoD, and IntoM stand for Objects in plate, Object into drawer, Object into microwave, respectively.

Figure 6: Mixed-Training Skill Interference and Continual-Learning Degradation. The top two rows illustrate skill interference in long-horizon tasks: the first shows successful single-skill executions, while the second shows failures after mixed training. The bottom two rows show degradation after continual learning: the first row presents the performance of $\pi_{0.5}$ before learning new skills, and the second shows its performance afterward. Red and green boxes highlight the key differences.

Method	Grasp	Stack	Close	Press	Open (new)	Avg.	$\Delta$ Avg.
$\pi_{0.5}$ [ 17 ]	85	65	70	90	-	77.5	–
$\pi_{0.5}$ [ 17 ] (CL)	70	45	60	75	55	61	- $15.0\downarrow$
AtomicVLA*	95	80	70	100	-	86.3	–
AtomicVLA* (CL)	90	80	80	100	70	82	$-1.3\downarrow$

Table 4: Continual Learning with Skill Expansion(%).

\Delta

Avg. represents the average performance change on the four base tasks after learning new skills compared with their performance before learning.CL is continual learning

Method	LIBERO-LONG
$\pi_{0}$ [ 3 ]	85.2
+ MoE	88.6
+ MoDE [ 39 ]	89.5
+ SG-MoE (Ours)	95.2

Table 5: Results on LIBERO Benchmark(%).

Atomic Skill	Count
Pick	2462
Place	761
Open	201
Close	152
Turn	175

Table 6: Atomic skill distribution in the LIBERO dataset.

Figure 7: The prompts and examples of the InternVideo2.5.

Task Type	Task Prompt
Long-horizon Tasks
Objects in plate	Place all blocks on the table into a green plate.
Object into drawer	Open the top drawer and place the block inside.
Object into microwave	Place the plate into the microwave and close the door.
Short Tasks
Grasp	Grasp the block from the table.

Table 7: List of tasks and prompts used in our real-world experiments.

Method	Pepper/Corn	Carrot/Cucumber	Potatoe/Eggplant	Avg.
$\pi_{0.5}$ [ 17 ]	25	40	35	33.3
AtomicVLA*	40	45	45	43.3

Table 8: Results on Complex Scenes.

Task Name	SR (%)	Task Name	SR (%)	Task Name	SR (%)
rotate blue block right	97.4	lift red block table	99.4	lift blue block table	99.4
move slider right	100.0	lift pink block table	94.5	place in drawer	100.0
lift red block slider	99.3	move slider left	100.0	rotate red block left	98.5
place in slider	98.6	turn on lightbulb	100.0	push pink block left	93.5
turn off lightbulb	100.0	rotate blue block left	100.0	lift blue block slider	95.6
turn off led	98.8	push blue block left	94.2	lift pink block drawer	100.0

Table 9: Success rates for all evaluated tasks on CALVIN ABC-D dataset.

Experts	$\pi_{0}$	K=5	K=8	K=12
Params	3.24B	4.17B	4.81B	5.65B
Act	71ms	92 ms	126 ms	160 ms
Think	-	104 ms	104 ms	104 ms

Table 10: Parameter and inference-time.

Figure 8: Demonstrations of LIBERO and Calvin experiments.

Figure 9: Error recovery cases in real-world experiments.

Figure 10: Demonstrations of real-world experiments(Long-horizon tasks).

Explore similar work

AtomVLA: Scalable Post-Training for Robotic Manipulation via Predictive Latent World Models
Mar 9, 2026Xiaoquan Sun, Zetian Xu, Chen Cao +9Vision-Language-Action Models Robotic Manipulation
INFIFORCE Intelligent Technology Co., Ltd. Hangzhou, China · Huazhong University of Science and Technology, Wuhan, China · The University of Hong Kong, Hong Kong SAR, China +1
LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies
Feb 25, 2026Yue Yang, Shuo Cheng, Yu Fang +4Long-Horizon Robotic Manipulation Robotic Manipulation
University of North Carolina at Chapel Hill · Georgia Institute of Technology · Carnegie Mellon University
Towards Human-like Physical Intelligence: LifelongVision-Language-Action Learning for Robotic Manipulation
Jul 16, 2026Yao He, Gan Sun, Wenqi Liang +2Continual Robot Learning Robot Skill Learning
South China University of Technology · University of Trento

Mar 9, 2026J/K move · Enter open · S save

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

Xiaoquan Sun, Zetian Xu, Chen Cao, Zonghe Liu+8

INFIFORCE Intelligent Technology Co., Ltd. Hangzhou, China · Huazhong University of Science and Technology, Wuhan, China · The University of Hong Kong, Hong Kong SAR, China+1

Vision-Language-Action Models Robotic Manipulation Long-Horizon Robotic Manipulation Robustness of VLA Models

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies

Towards Human-like Physical Intelligence: LifelongVision-Language-Action Learning for Robotic Manipulation