ABot-M0: VLA Foundation Model for Robotic Manipulation with Action Manifold Learning

Yandan Yang; Shuang Zeng; Tong Lin; Xinyuan Chang; Dekang Qi; Junjin Xiao; Haoyun Liu; Ronghan Chen; Yuzhi Chen; Dongjie Huo; Feng Xiong; Xing Wei; Zhiheng Ma; Mu Xu

Source-linked AI summary

ABot-M0: VLA Foundation Model for Robotic Manipulation with Action Manifold Learning

Yandan Yang, Shuang Zeng, Tong Lin, Xinyuan Chang, Dekang Qi, Junjin Xiao, Haoyun Liu, Ronghan Chen, Yuzhi Chen, Dongjie Huo, Feng Xiong, Xing Wei, Zhiheng Ma, Mu Xu

arXiv:2602.11236v2Feb 11, 2026cs.CVcs.CLcs.RO

TL;DR

General-purpose robotic agents must transfer across diverse hardware despite fragmented data, inconsistent representations, and mismatched training objectives. ABot-M0 unifies six public datasets and combines standardized pretraining with Action Manifold Learning and modular dual-stream perception. The model reports strong performance across four benchmarks, while future work targets data-capacity limits and finer control of task balance and embodiment coverage.

Problem
Robotic policies rarely transfer across platforms because embodied data are fragmented, representations differ, and existing VLA training objectives are mismatched.
Method
ABot-M0 curates and standardizes six open-source datasets, then combines unified pretraining with direct clean-action prediction and plug-and-play geometry-aware perception.
Results
98.6%, 80.5%, 58.3% and 81.2% average success rates are achieved on LIBERO, LIBERO-Plus, RoboCasa GR1 Tabletop Tasks and Robotwin2.0, respectively.
Takeaways & Limitations
Systematic engineering of public data, architecture, and training can support high-performance, generalizable embodied intelligence without proprietary data.
Takeaways & Limitations
Future work must address data-capacity limits and improve control over task balance and embodiment coverage.

Abstract

from arXiv · show

Building general-purpose embodied agents across diverse hardware remains a central challenge in robotics, often framed as the ''one-brain, many-forms'' paradigm. Progress is hindered by fragmented data, inconsistent representations, and misaligned training objectives. We present ABot-M0, a framework that builds a systematic data curation pipeline while jointly optimizing model architecture and training strategies, enabling end-to-end transformation of heterogeneous raw data into unified, efficient representations. From six public datasets, we clean, standardize, and balance samples to construct UniACT-dataset, a large-scale dataset with over 6 million trajectories and 9,500 hours of data, covering diverse robot morphologies and task scenarios. Unified pre-training improves knowledge transfer and generalization across platforms and tasks, supporting general-purpose embodied intelligence. To improve action prediction efficiency and stability, we propose the Action Manifold Hypothesis: effective robot actions lie not in the full high-dimensional space but on a low-dimensional, smooth manifold governed by physical laws and task constraints. Based on this, we introduce Action Manifold Learning (AML), which uses a DiT backbone to predict clean, continuous action sequences directly. This shifts learning from denoising to projection onto feasible manifolds, improving decoding speed and policy stability. ABot-M0 supports modular perception via a dual-stream mechanism that integrates VLM semantics with geometric priors and multi-view inputs from plug-and-play 3D modules such as VGGT and Qwen-Image-Edit, enhancing spatial understanding without modifying the backbone and mitigating standard VLM limitations in 3D reasoning. Experiments show components operate independently with additive benefits. We will release all code and pipelines for reproducibility and future research.

1 Introduction

ABot-M0 addresses fragmented robotic data, inconsistent representations, and mismatched action-learning objectives through unified data curation and architectural redesign. Its components collectively support cross-embodiment generalization, efficient action prediction, and strong benchmark performance.

Fragmented data, inconsistent action representations, and VLM-to-control mismatches hinder general-purpose policies across robot bodies.
ABot-M0 integrates six open-source datasets into UniACT-dataset, standardizing actions, coordinates, sampling, and task balance across 20+ embodiments.The resulting dataset contains over 6 million trajectories spanning 9500+ hours.
Action Manifold Learning predicts clean action sequences directly on a low-dimensional feasible manifold rather than learning unstructured noise.The approach is designed to improve decoding speed and policy stability.
A dual-stream perception design combines VLM semantic features with optional geometry-aware and multi-view modules such as VGGT and Qwen-Image-Edit.The plug-and-play perception stream compensates for standard VLM limitations in spatial understanding.
98.6%, 80.5%, 58.3% and 81.2% average success rates are reported on LIBERO, LIBERO-Plus, RoboCasa GR1 Tabletop Tasks and Robotwin2.0, respectively.The model outperforms strong baselines including π0.5, UniVLA, and OpenVLA-OFT.

2 Dataset

The dataset pipeline analyzes heterogeneous open-source VLA resources, cleans and standardizes their trajectories, and integrates them into a broad pretraining foundation. UniACT-dataset retains high-confidence data across diverse tasks and embodiments while aligning action and observation interfaces.

2.1 Analysis of Open-Source Datasets: Open-source VLA datasets differ in scale, quality, embodiment diversity, and representation, making it difficult for one dataset to satisfy all three priorities.
2.1 Analysis of Open-Source Datasets: Six datasets are integrated with dataset-specific curation strategies to balance large-scale coverage, data quality, and embodiment diversity.The sources include OXE, OXE-AugE, Agibot-Beta, RoboCoin, RoboMind, and Galaxea.
2.2 Data Cleaning and Preprocessing: All trajectories are converted to LeRobot v2, while invalid instructions, language inconsistencies, temporal misalignments, and visual anomalies are filtered or repaired.
2.2 Data Cleaning and Preprocessing: Approximately 16% of trajectories are discarded during cleaning, leaving more than six million trajectories across 9500+ hours and 20+ embodiments.The remaining high-confidence samples form UniACT-dataset for pretraining.
2.3 Standardization of Data Formats: The unified action interface uses end-effector delta actions and rotation vectors to improve training efficiency, rotational stability, and cross-embodiment compatibility.

3 The ABot-M0 Model

ABot-M0 combines a VLM with an action expert, using two-stage training and Action Manifold Learning to generate stable action sequences from heterogeneous multimodal inputs.

Model Architecture: ABot-M0 separates perception and action generation into a VLM and an action expert within a two-stage training framework.The VLM processes multi-view images and language, while the action expert produces executable action sequences.
Action Manifold Learning: Action Manifold Learning directly predicts denoised action chunks rather than noise or flow velocity targets.The action generator uses a DiT backbone and conditions on VLM or 3D features, robot state, and noisy actions.
Action Manifold Learning: Velocity-based training remains the optimization objective, because experiments found it superior to applying loss directly on actions.The estimated and ground-truth velocities are compared with a reweighted mean-squared-error loss.
Inference: During inference, iterative ODE integration updates actions from noisy initial states while retaining smooth and stable trajectory generation.The model predicts clean actions, derives flow velocity, and applies numerical integration at each timestep.
Two-Stage Training: UniACT pre-training uses approximately 6 million heterogeneous trajectories and standardized delta actions across single-arm and dual-arm tasks.Single-arm actions use seven dimensions, while dual-arm outputs use 14 dimensions; padding supports a shared network.
Two-Stage Training: Rare-task oversampling and batch-level configuration balancing reduce long-tail bias and improve zero-shot generalization on RoboCasa and Libero-Plus.The second stage then fine-tunes the model for spatially precise tasks while preserving performance on previously learned simple tasks.

4 Pre-Training

The pre-training pipeline addresses heterogeneous-data bias through curated sampling strategies and evaluates their effects on embodiment balance, skill coverage, generalization, and downstream transfer.

Sampling Motivation: Naive trajectory-level sampling can overrepresent frequent data, reducing exposure to rare skills and robot morphologies.The resulting distribution repeatedly samples frequent behaviors while limiting coverage of long-tail domains.
Sampling Strategies: Task-uniform sampling alleviates embodiment concentration by increasing exposure to multi-task, single-embodiment RoboCoin data.Trajectory-uniform sampling remains dominated by AgiBot-G1, while embodiment-uniform sampling introduces stronger high-frequency-skill bias.
Sampling Strategies: Task-uniform sampling produces a less concentrated skill distribution and faster coverage of unique skills than the alternatives.Its Lorenz curve is closer to equality, with a lower Gini coefficient and more efficient use of the sampling budget.
Generalization Evaluation: Task-uniform and embodiment-uniform sampling have similar overall cross-embodiment MAE, while task-uniform is lower on most RoboCoin embodiments.Trajectory-uniform performs substantially worse across nearly all embodiments.
Generalization Evaluation: Task-uniform sampling achieves the lowest MAE on OXE, AgiBot-Beta, and RoboCoin in cross-dataset evaluation.The validation protocol uses dataset-specific diversity-aligned sampling and mean absolute error after identical 50k-step training.
Downstream Transfer: Sampling-strategy differences persist in Libero Plus fine-tuning, whose downstream performance follows the overall MAE trend.Libero Plus is used only for downstream supervised fine-tuning, not pre-training.

5 Perception to Action

ABot-M0 transfers action-relevant VLM features to an action expert and supplements semantic representations with optional 3D information for spatially precise manipulation.

3D Information Injection: VLMs provide semantic understanding but lack precise geometric awareness needed for reachability and metric spatial reasoning.This limitation is especially important for precise manipulation tasks.
VLM Feature Interaction: Deep final-layer VLM features outperform shallow, intermediate, multi-layer, and action-query alternatives for policy learning.Raw VLM features directly outperform concatenation with action queries, and multi-layer aggregation adds no significant improvement.
3D Information Injection: A plug-and-play 3D module adds geometric priors alongside VLM semantics without replacing the VLM.VGGT supplies single-image 3D features, while Qwen-Image-Edit provides implicit multi-view features for robustness under occlusion and viewpoint changes.
3D Information Injection: Single-layer cross-attention provides the best fusion strategy for harmonizing semantic and geometric features before the action expert.The compared alternatives are concatenation, cross-attention, and Q-Former fusion.
3D Information Injection: Two synthesized views improve viewpoint-sensitive performance, including a reported +14% points on LIBERO-Plus camera perturbation tasks.The final system uses two synthetic views and two-step inference, fused through cross-attention.
3D Information Injection: The modular 3D stream can be enabled, disabled, or combined without retraining the core VLM.This supports task-dependent deployment of semantic and geometric processing.

6.1 Experiment Settings

ABot-M0 is evaluated across multiple simulation benchmarks to assess generalization, robustness, and adaptability across single-arm and dual-arm robots.

Experiment Settings: Experiments cover LIBERO, LIBERO-Plus, RoboCasa GR1 Tabletop Tasks, and Robotwin2.0.The evaluation uses Qwen3-VL 4B as the VLM backbone and a 0.16B DiT with Action Manifold Learning as the action expert.

6.2 Main Results

ABot-M0 achieves strong performance across LIBERO, LIBERO-Plus, RoboCasa GR1 Tabletop Tasks, and RoboTwin2.0, including zero-shot robustness and high-dimensional manipulation evaluation.

LIBERO: 98.6% average success rate on LIBERO demonstrates strong performance across jointly trained test suites.The model reaches 98.8% and 96.6% on the two spatial long-horizon trajectory tasks.
LIBERO-Plus: 80.5% success rate on LIBERO-Plus is achieved zero-shot after training only on standard LIBERO.This exceeds prior OpenVLA and OFT methods by 12.6–64.9%.
RoboCasa GR1 Tabletop Tasks: 58.3% success rate on RoboCasa GR1 Tabletop Tasks surpasses prior noise-prediction and action-regression experts.The evaluation covers 24 articulated-object manipulation tasks and predicts 464 high-dimensional actions.
RoboTwin2.0: Over 80% success rate on RoboTwin2.0 is achieved in both clean and heavily randomized multi-task settings.Training uses 2,500 clean-scene and 25,000 randomized-scene demonstrations.

6.3 Ablation Study

Ablations show that AML improves action-generation performance under typical and challenging settings, while feature-interaction and 3D modules provide additional gains.

Action Manifold Learning: 1.7% higher performance than GR00T under four denoising steps and action chunks of eight supports direct action prediction over noise prediction.The comparison controls initialization, action-expert size, and excludes an additional 3D spatial module.
Action Manifold Learning: 72.4 success rate is reached by ABot-M0 with action chunk size 10, while GR00T remains unchanged.At chunk size 30, GR00T drops by 23.6%, whereas ABot-M0 maintains 62.8%.
VLM Feature Interaction: 71% success rate is obtained when action queries use final-hidden-layer VLM features.The final hidden layer is reported as the most abstract and task-relevant feature source; 64 queries balance efficiency and performance.
3D Feature Injection: VGGT feature integration consistently improves both LIBERO and LIBERO-Plus, with single-layer Cross-Attention performing best.The ablation keeps the pretrained Qwen3-VL 4B model fixed and varies only the evaluated components.
3D Feature Injection: 2.8% absolute gain on LIBERO and 2–4% gains on LIBERO-Plus result from synthesized multi-view features.Using two synthesized views improves the camera-viewpoint-perturbation subset by up to 14%.

7 Future Work & Conclusion

ABot-M0 concludes by combining unified data standardization, action-manifold learning, and dual-stream 3D perception into a hardware-agnostic framework. Future work targets larger and better-balanced data, self-evolving curation, multimodal sensing, and intrinsic geometric modeling.

Future Work: Future work will scale human and UMI-collected demonstrations while studying finer control of data quality, task balance, and embodiment coverage.The authors identify data capacity and curation quality as areas requiring further development.
Future Work: Future directions include self-evolving data curation, unified multimodal sensing, and intrinsic 3D representation learning.The proposed directions connect execution feedback, additional sensing modalities, and self-supervised geometric priors.
Conclusion: ABot-M0 integrates six open-source datasets into UniACT-dataset and applies standardized action representations to support cross-embodiment transfer.The framework uses 6M+ trajectories, end-effector-frame delta actions, pad-to-dual-arm modeling, and dual-level reweighting.
Conclusion: A plug-and-play dual-stream architecture fuses VLM semantics with enhanced 3D perception without requiring replacement of the VLM.The 3D module adds geometric priors alongside the VLM to support reasoning about what to do and where to act.

8 Contributions

The listed contributions span data collection and standardization, pipeline development, model architecture, training, evaluation, writing, project leadership, and advising.

Contributions: The contribution list assigns separate responsibilities for data collection and analysis, data standardization, and data-pipeline development.These roles are attributed to named contributors in the author-contribution statement.
Contributions: Model architecture, training, and evaluation are identified as distinct contribution areas.The statement lists named contributors for each of these technical activities.
Contributions: Writing, project leadership, and advising are also explicitly included in the contribution statement.The listed roles cover manuscript preparation, project direction, and advisory support.