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

ABot-M0

ABot-M0 is a general-purpose robotics model with the following core highlights:

The overview of ABot-M0.

Massive & Unified Data: It integrates over 6 million open-source trajectories to form the largest unified dataset for robotic manipulation, providing a strong foundation for generalization.
Innovative Action Paradigm: It pioneers Action Manifold Learning (AML), which directly predicts clean actions instead of noise, resulting in a more efficient and stable model.
Modular 3D Perception: It supports plug-and-play modules to enhance 3D spatial understanding, improving execution precision for complex tasks.

UniACT-dataset

UniACT-dataset: a large-scale unified robotic manipulation dataset behind ABot-M0.

Robotic data suffers from fragmentation and inconsistent representations. To address this challenge, we constructed UniACT-dataset, one of the largest non-private robotic manipulation datasets to date.

Data Sources and Scale

Integration: It aggregates 6 major public datasets, including OXE, OXE-AugE, and AgiBot-Beta.
Scale: The dataset comprises over 6 million trajectories, amounting to 9,500+ hours of interaction data.
Diversity: It covers more than 20 different robot morphologies.

Systematic Data Curation Pipeline

Cleaning: We filter out data with invalid instructions, visual anomalies, and erroneous actions.
Standardization: All actions are unified into delta actions in the end-effector coordinate frame, with rotations adopting the more stable rotation vector representation.
Unification: A “pad-to-dual” strategy is employed to enable a single model to handle both single-arm and dual-arm tasks.

Action Manifold Learning (AML)

Action Manifold Learning (AML): learning robot actions on a low-dimensional, structured manifold.

Conventional diffusion models typically learn to predict noise, an approach that is inefficient and unstable for robotic control. We propose the Action Manifold Hypothesis: effective robot actions are not randomly distributed in a high-dimensional space, but rather lie on a low-dimensional, smooth manifold shaped by physical laws and task constraints.

Building on this, we design Action Manifold Learning (AML):

Direct Action Prediction (a-prediction): The model directly outputs clean action sequences instead of high-dimensional, unstructured noise.
Improved Efficiency and Stability: The learning objective shifts from “fitting noise” to “projecting onto the feasible manifold.” This enables the model to learn meaningful action structures more efficiently, thereby boosting decoding speed and policy stability.
Enhanced Scalability: The advantages of AML become more pronounced when handling higher-dimensional actions (e.g., for dexterous hands or whole-body control), laying a foundation for the long-term development of embodied intelligence.

Results

Success rates (%) on the LIBERO benchmark. ABot-M0 is one policy for all 4 suites.
Method	L-Spatial	L-Object	L-Goal	L-Long	Average
Diffusion Policy	78.5	87.5	73.5	64.8	76.1
OpenVLA	84.7	88.4	79.2	53.7	76.5
SpatialVLA	88.2	89.9	78.6	55.5	78.1
CoT-VLA	87.5	91.6	87.6	69.0	83.9
π₀-Fast	96.4	96.8	88.6	60.2	85.5
GR00T-N1	94.4	97.6	93.0	90.6	93.9
π₀	98.0	96.8	94.4	88.4	94.4
F1	98.2	97.8	95.4	91.3	95.7
InternVLA-M1	98.0	99.0	93.8	92.6	95.9
Discrete Diffusion VLA	97.2	98.6	97.4	92.0	96.3
π₀.₅	98.8	98.2	98.0	92.4	96.9
GR00T-N1.6	97.7	98.5	97.5	94.4	97.0
OpenVLA-OFT	97.6	98.4	97.9	94.5	97.1
X-VLA	98.2	98.6	97.8	97.6	98.1
ABot-M0 (Ours)	98.8	99.8	99.0	96.6	98.6

Zero-shot performance comparison on LIBERO-Plus. One ABot-M0 for all 4 suites is trained only on the standard LIBERO dataset (no LIBERO-Plus data fine-tuning).
Method	Camera	Robot	Language	Light	Background	Noise	Layout	Total
OpenVLA	0.8	3.5	23.0	8.1	34.8	15.2	28.5	15.6
OpenVLA-OFT	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
OpenVLA-OFT_w	10.4	38.7	70.5	76.8	93.6	49.9	69.9	55.8
NORA	2.2	37.0	65.1	45.7	58.6	12.8	62.1	39.0
WorldVLA	0.1	27.9	41.6	43.7	17.1	10.9	38.0	25.0
UniVLA	1.8	46.2	69.6	69.0	81.0	21.2	31.9	42.9
π₀	13.8	6.0	58.8	85.0	81.4	79.0	68.9	53.6
π₀-Fast	65.1	21.6	61.0	73.2	73.2	74.4	68.8	61.6
RIPT-VLA	55.2	31.2	77.6	88.4	91.6	73.5	74.2	68.4
Openvla-OFT_m	55.6	21.7	81.0	92.7	91.0	78.6	68.7	67.9
ABot-M0 (Ours)	60.4	67.9	86.4	96.2	91.6	86.4	82.6	80.5

RoboCasa GR1 Tabletop Tasks Evaluation Results. A single model was trained for all 24 tasks. Results are reported over 50 rollouts per task.
Task	GR00T-N1.6	Qwen3GR00T	Qwen3PI	Qwen3OFT	Qwen3FAST	ABot-M0 (Ours)
PnPBottleToCabinetClose	51.5	46.0	26.0	30.0	38.0	86.0
PnPCanToDrawerClose	13.0	80.0	62.0	76.0	44.0	74.0
PnPCupToDrawerClose	8.5	54.0	42.0	44.0	56.0	48.0
PnPMilkToMicrowaveClose	14.0	48.0	50.0	44.0	44.0	46.0
PnPPotatoToMicrowaveClose	41.5	28.0	42.0	32.0	14.0	50.0
PnPWineToCabinetClose	16.5	46.0	32.0	36.0	14.0	66.0
PnPNovelFromCuttingboardToBasket	58.0	48.0	40.0	50.0	54.0	70.0
PnPNovelFromCuttingboardToCardboardbox	46.5	40.0	46.0	40.0	42.0	58.0
PnPNovelFromCuttingboardToPan	68.5	68.0	60.0	70.0	58.0	76.0
PnPNovelFromCuttingboardToPot	65.0	52.0	40.0	54.0	58.0	66.0
PnPNovelFromCuttingboardToTieredbasket	46.5	56.0	44.0	38.0	40.0	38.0
PnPNovelFromPlacematToBasket	58.5	42.0	44.0	32.0	36.0	52.0
PnPNovelFromPlacematToBowl	57.5	44.0	52.0	58.0	38.0	66.0
PnPNovelFromPlacematToPlate	63.0	48.0	50.0	52.0	42.0	60.0
PnPNovelFromPlacematToTieredshelf	28.5	18.0	28.0	24.0	18.0	26.0
PnPNovelFromPlateToBowl	57.0	60.0	52.0	60.0	52.0	54.0
PnPNovelFromPlateToCardboardbox	43.5	50.0	40.0	50.0	30.0	48.0
PnPNovelFromPlateToPan	51.0	54.0	36.0	66.0	48.0	66.0
PnPNovelFromPlateToPlate	78.7	70.0	48.0	68.0	50.0	64.0
PnPNovelFromTrayToCardboardbox	51.5	38.0	34.0	44.0	28.0	54.0
PnPNovelFromTrayToPlate	71.0	56.0	64.0	56.0	34.0	68.0
PnPNovelFromTrayToPot	64.5	50.0	44.0	62.0	46.0	64.0
PnPNovelFromTrayToTieredbasket	57.0	36.0	50.0	54.0	36.0	60.0
PnPNovelFromTrayToTieredshelf	31.5	16.0	28.0	30.0	16.0	38.0
Average	47.6	47.8	43.9	48.8	39.0	58.3

Evaluation on RoboTwin 2.0 Simulation (Clean vs Randomized, 50+ tasks).
Simulation Task	π₀.₅	X-VLA	ABot-M0 (Ours)
Place Dual Shoes	12	7	79	88	75	79
Move Stapler Pad	16	18	78	73	49	72
Stack Blocks Two	48	56	92	87	100	98
Scan Object	42	38	14	36	100	80
Place Object Stand	74	65	86	88	100	94
Place Fan	25	36	80	75	84	91
Move Pillbottle Pad	33	29	73	71	93	100
Pick Dual Bottles	10	6	47	36	77	100
Blocks Ranking Rgb	43	35	83	83	68	85
......(50 tasks)	...	...	...	...	...	...
Turn Switch	5	6	40	61	52	81
Pick Diverse Bottles	5	3	58	36	70	93
Place Bread Basket	48	56	81	71	91	87
Stack Blocks Three	15	16	6	10	72	49
Put Bottles Dustbin	12	9	74	77	74	70
Place Can Basket	19	25	49	52	45	66
Stamp Seal	36	23	76	82	76	76
Handover Block	18	19	73	37	93	69
Stack Bowls Three	33	35	76	86	100	88
Place Object Basket	43	36	44	39	97	95
Open Microwave	35	37	79	71	93	49
Average	42.98	43.84	72.80	72.84	80.42	81.16

Video

We showcase representative rollouts from the four benchmark suites used to evaluate ABot-M0.

LIBERO

LIBERO-Plus

RoboCasa

RoboTwin 2.0

Author Team

Author contributions in the following areas are as follows:

Data Collection & Analysis: Yandan Yang, Haoyun Liu, Ronghan Chen, Yuzhi Chen, Dekang Qi
Data Standardization: Yandan Yang, Tong lin, Xinyuan Chang
Data Pipeline: Tong Lin, Shuang Zeng, Dongjie Huo
Model Architecture: Shuang Zeng, Junjin Xiao
Training: Shuang Zeng, Tong Lin, Junjin Xiao
Evaluation: Junjin Xiao, Shuang Zeng, Tong Lin
Project Lead: Xinyuan Chang, Feng Xiong
Advisor: Mu Xu†, Zhiheng Ma, Xing Wei

Citation

If you find our work helpful, please cite us:


              @article{abot-m0,
                title={ABot-M0: VLA Foundation Model for Robotic Manipulation with Action Manifold Learning},
                author={AMAP CV Lab},
                year={2026}
              }

Thank you!

Simulation Task	π₀.₅		X-VLA		ABot-M0 (Ours)
	Clean	Rand.	Clean	Rand.	Clean	Rand.
Place Dual Shoes	12	7	79	88	75	79
Move Stapler Pad	16	18	78	73	49	72
Stack Blocks Two	48	56	92	87	100	98
Scan Object	42	38	14	36	100	80
Place Object Stand	74	65	86	88	100	94
Place Fan	25	36	80	75	84	91
Move Pillbottle Pad	33	29	73	71	93	100
Pick Dual Bottles	10	6	47	36	77	100
Blocks Ranking Rgb	43	35	83	83	68	85
......(50 tasks)	...	...	...	...	...	...
Turn Switch	5	6	40	61	52	81
Pick Diverse Bottles	5	3	58	36	70	93
Place Bread Basket	48	56	81	71	91	87
Stack Blocks Three	15	16	6	10	72	49
Put Bottles Dustbin	12	9	74	77	74	70
Place Can Basket	19	25	49	52	45	66
Stamp Seal	36	23	76	82	76	76
Handover Block	18	19	73	37	93	69
Stack Bowls Three	33	35	76	86	100	88
Place Object Basket	43	36	44	39	97	95
Open Microwave	35	37	79	71	93	49
Average	42.98	43.84	72.80	72.84	80.42	81.16