SLA: Scaling Spatial-Language-Action Model via Continual Pre-training

Fuhao Li; Wenxuan Song; Qizhou Huang; Ziyang Zhou; Jingbo Wang; Junwu E; Pengwei Wang; Yan Wang

cs.RONov 5, 2026

SLA: Scaling Spatial-Language-Action Model via Continual Pre-training

Authors: Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou, Jingbo Wang, Junwu E, Pengwei Wang, Yan Wang,

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · Beijing Academy of Artificial Intelligence · Tsinghua University · OLA Dimensions

Abstract

Reliable robotic manipulation requires both semantic understanding and precise control in three-dimensional space. However, vision-language-action (VLA) foundations largely inherit visual representations from 2D vision-language pre-training, while spatial supervision is often introduced only during task-specific adaptation. Developing spatial understanding as a transferable foundation-model capability therefore calls for learning from broader robot experience. We present SLA, a Spatial-Language-Action Model that scales implicit spatial alignment to continual pre-training and maintains it during downstream post-training. Initialized from a pretrained pi05 policy, SLA jointly optimizes action prediction and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model. We curate and unify heterogeneous robot datasets to support spatial learning across diverse embodiments, tasks, and scenes. To accommodate this diversity, we introduce feature-dependent gates that adaptively aggregate visual embeddings from multiple VLA layers for spatial supervision. The teacher further exploits partially available depth and camera calibration alongside RGB observations, enabling effective use of heterogeneous geometric annotations. Probabilistic state dropout encourages reliance on visual cues and mitigates overfitting to demonstration trajectories. All auxiliary components are used only during training, allowing direct reuse of pretrained VLA weights and preserving the original inference architecture.

Figures & tables

Figure 1: SLA unifies heterogeneous robot data, aligns intermediate visual embeddings with frozen 3D features during training, and improves benchmark and real-world manipulation performance.

Figure 2: Depth probing of the visual embeddings of VLAs. Embeddings learned solely from 2D images without alignment do not produce meaningful spatial structures. The aligned embeddings inherently contain rich spatial information, leading to better performance in depth probing.

Data source	Episodes	Tasks	Hours
InternData-A1 ( Tian et al., 2025 )	630K	70	7430
AgiBotWorld-Beta ( AgiBot-World-Contributors et al., 2025 )	168K	217	2970
RoboCasa365-Pretrain ( Nasiriany et al., 2026 )	630K	300	2090
RoboCOIN ( Wu et al., 2025b )	183K	808	1300
DROID ( Khazatsky et al., 2024 )	95K	86	510
MolmoAct ( Lee et al., 2026 )	7.9K	23	30

Table 1: Overview of the source datasets for continual pre-training. Statistics refer to the source datasets before our data selection and processing; task counts are summed across datasets.

Figure 3: Comparison of rotational action distributions. (a) Euler-angle representations with direct subtraction for relative-rotation targets. (b) Rotation-vector representations with relative rotations computed through $\mathrm{SO}(3)$ composition. In (a), angular wrapping produces outliers near $\pm 2\pi$ , which widen the interval between the 1st and 99th percentiles. This severely compresses the range of typical action values after normalization, hindering training.

Method	Camera	Robot	Language	Light	Background	Noise	Layout	Average
$\pi_{0}$ ( Black et al., 2025 )	13.8	6.0	58.8	85.0	81.4	79.0	68.9	53.6
$\pi_{0.5}$ ( Intelligence, 2025 )	78.4	73.6	80.8	96.2	94.1	89.0	84.5	84.4
StarVLA ( Community, 2026 )	52.5	49.8	88.5	95.7	95.7	73.0	76.9	74.1
4D-WAM ( Yang et al., 2026a )	45.1	64.3	90.6	94.3	57.7	69.1	79.2	71.0
Abot-M0 ( Yang et al., 2026c )	60.4	67.9	86.4	96.2	91.6	86.4	82.6	80.5
Cosmos Policy ( Kim et al., 2026 )	75.8	63.3	81.7	96.5	88.9	92.7	82.2	82.2

Table 2: LIBERO-Plus benchmark success rates (%,

\uparrow

). Models are fine-tuned on the standard LIBERO training set without LIBERO-Plus set. Average is weighted by the number of evaluation episodes. Bold and underline denote the best and second-best displayed values, respectively.

Method	Clean2Clean	Clean2Random	Average
Discrete Forcing ( Wang et al., 2026c )	59.3	11.2	35.3
X-WAM ( Guo et al., 2026a )	70.0	25.8	47.9
$\pi_{0.5}$ ( Intelligence, 2025 )	70.7	46.0	58.3
4D-WAM ( Yang et al., 2026a )	81.5	41.8	61.7
WorldScape Policy 2.0 ( Su et al., 2026a )	84.9	38.5	61.7
GigaBrain-0.7 ( Team et al., 2026a )	66.8	67.9	67.4

Table 3: RoboTwin 2.0 benchmark success rates (%,

\uparrow

). Models are fine-tuned jointly on the 50 standard tasks using 50 demonstrations from the Clean setting per task. Per-task results are shown in Table 12 . Bold and underline denote the best and second-best displayed values, with ties treated equally.

Method	Training Demos per Task	Average
GR00T-N1 ( Bjorck et al., 2025 )	300	49.6
UVA ( Li et al., 2025b )	50	50.0
DP-VLA ( Han et al., 2024 )	3000	57.3
DreamGen ( Jang et al., 2025 )	300	57.6
DUST ( Won et al., 2025 )	300	58.5
UWM ( Zhu et al., 2025 )	1000	60.8

Table 4: RoboCasa-Panda benchmark success rates (%,

\uparrow

). Per-task results are shown in Table 13 . Bold and underline denote the best and second-best displayed values.

Method	ARX X5		Piper		Piper X		Average
Method	Score	SR	Score	SR	Score	SR	Score	SR
Spirit v1.5 ( Team, 2026 )	0.0	0.0	4.0	1.7	0.7	0.0	1.6	0.6
StarVLA ( Community, 2026 )	0.0	0.0	12.0	5.0	0.5	0.0	4.1	1.7
$\pi_{0}$ ( Black et al., 2025 )	8.1	0.0	9.0	5.0	0.3	0.0	5.8	1.7
GR00T-N1.7 ( Bjorck et al., 2025 )	9.3	1.7	6.8	3.3	1.4	0.0	5.9	1.7
X-VLA ( Zheng et al., 2025a )	6.8	1.7	15.9	8.3	0.1	0.0	7.6	3.3

Table 5: RoboDojo real-world benchmark results across three embodiments. Score (

\uparrow

) captures partial task progress, and SR (%,

\uparrow

) measures binary task completion. Per-task results are shown in Table 14 . Bold and underline denote the best and second-best displayed values.

SLA Alignment		Continual Pre-training Data (%)
Pre-training	Post-training	5	25	50	100
✗	✗	43.8	46.5	46.1	48.2
✗	✓	45.6	53.4	55.3	57.1
✓	✗	46.6	52.0	58.4	63.2
✓	✓	49.2	60.5	67.4	72.3

Table 6: Continual pre-training data scaling ablation. SLA alignment is independently enabled or disabled during continual pre-training and post-training. The corresponding line chart is plotted in Figure 4 . Experiments are conducted under the RoboTwin Clean2Random settings to evaluate success rates (%,

\uparrow

Figure 4: Continual pre-training data scaling ablation. SLA alignment is independently enabled or disabled during continual pre-training and post-training. All curves start at the shared $\pi_{0.5}$ baseline. Experiments are conducted under RoboTwin Clean2Random settings to evaluate success rates (%, $\uparrow$ ).

Alignment Layer	Clean2Clean	Clean2Random	Average
1	60.0	43.4	51.7
6	63.0	45.0	54.0
12	84.0	67.3	75.7
18	78.0	64.6	71.3
Adaptive	86.0	72.3	79.2

Table 7: Adaptive layer selection ablation. Fixed variants use visual tokens from a single VLA transformer layer for alignment; Adaptive variants use the learned fusion mechanism. Experiments are conducted on RoboTwin to evaluate success rates (%,

\uparrow

State Dropout Probability	Clean2Clean	Clean2Random	Average
0%	55.5	20.7	38.1
25%	86.0	72.3	79.2
50%	66.5	41.3	53.9
75%	72.1	51.6	61.9
100%	62.4	50.9	56.7

Table 8: State dropout ablation. Experiments are conducted on RoboTwin to evaluate success rates (%,

\uparrow

Rotation Representation	Clean2Clean	Clean2Random	Average
Euler Angle	71.6	59.3	65.5
Rot6D	66.9	58.8	62.9
Quaternion	81.1	67.5	74.3
Rotation Vector	86.0	72.3	79.2

Table 9: Rotation representation ablation. All variants compute relative rotations through

\mathrm{SO}(3)

composition rather than direct subtraction to avoid wrap-induced outliers, as detailed in Equation 10 . Experiments are conducted on RoboTwin to evaluate success rates (%,

\uparrow

SLA Alignment	Clean2Clean	Clean2Random	Average
✗	36.9	26.7	31.8
VGGT	50.2	34.8	42.5
OmniVGGT w/o Calibration	49.9	35.6	42.8
OmniVGGT w/ Calibration	54.7	37.3	46.0

Table 10: Camera calibration ablation. Calibration comprises camera intrinsics and extrinsics supplied to the OmniVGGT for spatial supervision. Experiments are conducted on RoboTwin to evaluate success rates (%,

\uparrow

). Models are fine-tuned on personal-collected RoboTwin clean data.

Figure 5: Produce manipulation tasks under human perturbations and embodiment changes. A–C conditions share a policy post-trained on 100 demonstrations collected in the A scene configuration, with no additional demonstrations of the perturbations in B or C scene. D condition evaluates a policy jointly post-trained on 100 Piper demonstrations collected in the A scene and 10 Franka demonstrations collected in the D scene.

Figure 6: Tabletop manipulation tasks under changes in object, table height, and instruction target. All four conditions share a policy post-trained on 100 demonstrations collected in the E scene configuration. The bottle, tissue box, pen, and watch are included in these demonstrations while black tape and carrot are excluded.

Method	A: Pass vegetables between hands	B: Grasp fruit thrown by a human	C: Grasp fruit moved by a human	D: Another robot grasps vegetables
OpenVLA-OFT ( Kim et al., 2025 )	62	38	30	36
SpatialVLA ( Qu et al., 2025 )	70	46	46	44
$\pi_{0.5}$ ( Intelligence, 2025 )	74	52	42	48
Spatial Forcing	84	66	58	62
SLA (Ours)	94	82	76	78

Table 11: Real-world deployment comparisons. Each method is evaluated on 50 separately scored object instances per task to report the object-level success rates (%,

\uparrow

). Average is the arithmetic mean over A–H. Bold and underline denote the best and second-best displayed values.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Clean2Clean	Clean2Random	Task	Clean2Clean	Clean2Random
Adjust Bottle	99	99	Place Can Basket	80	57
Beat Block Hammer	83	68	Place Cans Plasticbox	96	40
Blocks Ranking RGB	94	90	Place Container Plate	100	98
Blocks Ranking Size	68	54	Place Dual Shoes	61	41
Click Alarmclock	100	96	Place Empty Cup	99	94
Click Bell	100	93	Place Fan	75	66

Appendix

Table 12: Per-task success rates of SLA on RoboTwin 2.0 (%,

\uparrow

). Each task is evaluated for 100 trials per setting. Task order follows the official leaderboard ( Chen et al., 2025b ) : tasks 1–25 are on the left and 26–50 on the right.

Task	Mean SR (%) $\uparrow$	Task	Mean SR (%) $\uparrow$
PnPCounterToCab	54.7	OpenDrawer	84.0
PnPCabToCounter	21.3	CloseDrawer	97.3
PnPCounterToSink	68.0	TurnOnStove	49.3
PnPSinkToCounter	58.0	TurnOffStove	22.7
PnPCounterToMicrowave	41.3	TurnOnSinkFaucet	72.0
PnPMicrowaveToCounter	34.7	TurnOffSinkFaucet	94.0

Appendix

Table 13: Per-task success rates of SLA on RoboCasa-Panda (%,

\uparrow

). Each value is the arithmetic mean over three evaluation seeds (195, 196, and 197), with 50 trials per task and seed (150 trials per task; 3,600 trials in total). Tasks follow the evaluation-protocol order, with tasks 1–12 on the left and 13–24 on the right.

Embodiment	Task	Score $\uparrow$	SR (%) $\uparrow$
ARX X5	cover_blocks	65.0	50
	store_in_safe	100.0	100
	pack_and_pour_fruit	70.0	40
	make_bread	0.0	0
	make_food	29.0	20
	insert_tubes	40.0	20

Appendix

Table 14: Per-task results of SLA on the RoboDojo real-world benchmark. Score (

\uparrow

) is scaled to 0–100, and SR (%,

\uparrow

) measures task completion. Means give equal weight to the six tasks within each embodiment; the overall average gives equal weight to the three embodiments.

Method	Generalization		Precision		Long-Horizon		Memory		Open		Average
Method	Score	SR	Score	SR	Score	SR	Score	SR	Score	SR	Score	SR
InternVLA-A1.5 ( Cai et al., 2026a )	10.4	6.8	15.2	10.2	23.8	13.8	4.9	3.6	1.4	1.4	11.2	7.1
$\pi_{0.5}$ ( Intelligence, 2025 )	13.4	8.2	12.4	5.5	23.5	14.7	5.9	4.7	2.0	1.7	11.4	6.9
SimpleMemVLA ( Yin et al., 2026 )	6.4	4.0	7.4	2.9	14.6	5.5	33.7	33.2	0.9	0.8	12.6	9.3
KinRT ( Yang et al., 2026b )	14.0	8.6	15.7	9.9	26.4	18.1	4.8	3.6	4.2	3.8	13.0	8.8
CogWAM ( Wang et al., 2026d )	12.8	9.3	20.5	14.6	27.4	18.2	8.2	7.0	2.9	2.7	14.3	10.3

Appendix

Table 15: RoboDojo simulation benchmark results across five capability dimensions. Score (

\uparrow

) captures partial task progress, and SR (%,

\uparrow

) measures binary task completion. Results of the compared methods are taken from the official leaderboard ( Chen et al., 2026b ) , and SLA is evaluated locally. Per-task results are shown in Table 16 . Bold and underline denote the best and second-best displayed values.

Task	Score $\uparrow$	SR (%) $\uparrow$	Task	Score $\uparrow$	SR (%) $\uparrow$
stack_bowls	74.5	70	pour_balls_into_vase	20.0	20
stack_bowls_random	19.0	10	play_Xylophone	0.0	0
push_T	0.0	0	deposit_coin	2.0	0
push_T_random	0.0	0	insert_key	12.8	0
pack_objects_into_box	17.2	5	build_tower	69.0	65
pack_objects_into_box_random	14.8	0	put_bottles_into_dustbin	76.8	65

Appendix

Table 16: Per-task results of SLA on the RoboDojo simulation benchmark. Score (

\uparrow

) is scaled to 0–100, and SR (%,

\uparrow

) measures task completion, as in Table 15 . Task order follows the official leaderboard ( Chen et al., 2026b ) , with each Generalization task followed by its randomized (_random) variant; configurations 1–27 are on the left and 28–54 on the right. The overall average gives equal weight to the five capability dimensions.

VLA Model Training

See all 54 papers

Universal Pose Pretraining for Generalizable Vision-Language-Action Policies
Feb 23, 2026Haitao Lin, Hanyang Yu, Jingshun Huang +5Robot Policy Learning Vision-Language-Action Models
Tencent Robotics X · Futian Laboratory · The Hong Kong University of Science and Technology +2
Relational Spatial Alignment for Vision-Language-Action Models
Oct 17, 2026Chaoran Zhu, Hengyi Wang, Chenglin Cui +3Representation Alignment 3D Spatial Reasoning
Queen Mary University of London, United Kingdom · University College London, United Kingdom · VGG, University of Oxford, United Kingdom
VEGA: Visual Encoder Grounding Alignment for Spatially-Aware Vision-Language-Action Models
May 11, 2026Hao Wang, Xiaobao Wei, Jingyang He +103D Spatial Reasoning Spatial Grounding for VLA Models
1Peking University, Beijing, China · 2Beijing Innovation Center of Humanoid Robotics, Beijing, China

Feb 23, 2026J/K move · Enter open · S save

Universal Pose Pretraining for Generalizable Vision-Language-Action Policies

Haitao Lin, Hanyang Yu, Jingshun Huang, He Zhang+4

Tencent Robotics X · Futian Laboratory · The Hong Kong University of Science and Technology+2

Robot Policy Learning Vision-Language-Action Models Robot Imitation Learning 3D Representation Learning Robot Policy Generalization

Open paper AskPDF arXiv

Existing Vision-Language-Action (VLA) models often suffer from feature collapse and low training efficiency because they entangle high-level perception with sparse, embodiment-specific action supervision. Since these models typically rely on VLM backbones optimized for Visual Question Answering (VQA), they excel at semantic identification but often overlook subtle 3D state variations that dictate distinct action patterns. To resolve these misalignments, we propose Pose-VLA, a decoupled paradigm that separates VLA training into a pre-training phase for extracting universal 3D spatial priors in a unified camera-centric space, and a post-training phase for efficient embodiment alignment within robot-specific action space. By introducing discrete pose tokens as a universal representation, Pose-VLA seamlessly integrates spatial grounding from diverse 3D datasets with geometry-level trajectories from robotic demonstrations. Our framework follows a two-stage pre-training pipeline, establishing fundamental spatial grounding via poses followed by motion alignment through trajectory supervision. Extensive evaluations demonstrate that Pose-VLA achieves state-of-the-art results on RoboTwin 2.0 with a 79.5% average success rate and competitive performance on LIBERO at 96.0%. Real-world experiments further showcase robust generalization across diverse objects using only 100 demonstrations per task, validating the efficiency of our pre-training paradigm.

Abstract

VLA Model Training

Explore similar work

Universal Pose Pretraining for Generalizable Vision-Language-Action Policies

Relational Spatial Alignment for Vision-Language-Action Models

VEGA: Visual Encoder Grounding Alignment for Spatially-Aware Vision-Language-Action Models