Source-linked AI summary
SimVLA: A Simple VLA Baseline for Robotic Manipulation
Yuankai Luo, Woping Chen, Tong Liang, Baiqiao Wang, Zhenguo Li
TL;DR
VLA progress has made it difficult to separate architectural gains from differences in training recipes and implementation details. SimVLA addresses this with a modular perception-control split and standardized training dynamics, achieving state-of-the-art benchmark performance with a compact model. The authors present it as a reproducible reference for evaluating future VLA innovations, while cautioning that real-world deployment requires safety constraints and validation.
Problem
Varying architectures, datasets, backbone scales, and optimization schedules make it difficult to disentangle the source of empirical VLA performance gains.
Method
SimVLA uses a standard VLM encoder, lightweight continuous-action head with flow matching, and standardized training and inference dynamics.
Results
SimVLA achieves state-of-the-art performance, including the highest reported average LIBERO success rate with a compact 0.5B backbone surpassing larger baselines.
Takeaways & Limitations
SimVLA provides a reproducible reference point for more precise attribution of gains from future VLA architectural innovations.
Takeaways & Limitations
SimVLA is evaluated in controlled research settings, so real-world deployment requires safety constraints, human oversight, and domain-specific validation.
Abstract
from arXiv · showhide
Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robotic manipulation, leveraging large-scale pre-training to achieve strong performance. The field has rapidly evolved with additional spatial priors and diverse architectural innovations. However, these advancements are often accompanied by varying training recipes and implementation details, which can make it challenging to disentangle the precise source of empirical gains. In this work, we introduce SimVLA, a streamlined baseline designed to establish a transparent reference point for VLA research. By strictly decoupling perception from control, using a standard vision-language backbone and a lightweight action head, and standardizing critical training dynamics, we demonstrate that a minimal design can achieve state-of-the-art performance. Despite having only 0.5B parameters, SimVLA outperforms multi-billion-parameter models on standard simulation benchmarks without robot pretraining. SimVLA also reaches on-par real-robot performance compared to pi0.5. Our results establish SimVLA as a robust, reproducible baseline that enables clear attribution of empirical gains to future architectural innovations. Website: https://frontierrobo.github.io/SimVLA
1. Introduction
VLA research has added increasingly sophisticated architectural components, but confounded training recipes and implementation choices make performance gains difficult to attribute. SimVLA addresses this with a minimal, modular baseline that standardizes training and achieves strong performance with efficient deployment.
- Architectural innovations, varying pretraining, backbone scales, and optimization schedules complicate attribution of performance gains to specific VLA components.
- SimVLA provides a transparent reference point by decoupling perception from control while preserving a high-performance lower-complexity baseline.
- Standardized data shuffling, action normalization, and optimization schedules are treated as essential factors for fair cross-model comparisons.
- SimVLA outperforms multi-billion-parameter baselines while maintaining a compact memory footprint in a representative LIBERO evaluation.
- The baseline combines modular perception-control separation, standardized training dynamics, state-of-the-art performance, and efficient zero-shot real-robot transfer.
2. Related Work
Recent VLA work improves robotic control through visual-temporal augmentation, geometric priors, and richer action architectures. These additions can also increase inference latency or training instability, motivating SimVLA’s simpler flow-matching baseline.
- Visual and Temporal Augmentation: Recent VLA methods augment visual inputs with motion cues, predictive world models, planning or intention descriptions, and memory mechanisms.
- Geometric and 3D Priors: Other approaches inject 3D coordinates, egocentric positional encodings, depth-aware tokens, point clouds, geometric tokens, or depth maps for spatial understanding.
- Action Representations and Architectures: Action-representation research pursues computational efficiency through trajectory compression, parallel decoding, continuous regression, and lightweight adapters or action heads.
- Action Representations and Architectures: Other models represent continuous multimodal distributions with diffusion policies, unified tokenization, latent action spaces, or universal discrete action codebooks.
- Despite reported gains, increasingly complex models can introduce inference latency or training instability; SimVLA instead uses standard flow matching and a validated training recipe.
3. SimVLA: A Simple VLA Baseline
SimVLA deliberately separates vision-language perception from continuous action generation through a lightweight action head and flow matching. Its recipe controls normalization, chunking, shuffling, optimization, architecture, and efficient inference to provide a neutral baseline.
- SimVLA responds to confounded comparisons by using a simple modular baseline that decouples perception and control.
- Architecture: A pretrained VLM maps multi-view RGB observations and language into shared tokens, while the action head combines them with proprioception, timestep embeddings, and noised action chunks.
- Flow Matching: Flow matching constructs noised normalized action chunks from Gaussian noise and trains the action head to predict the denoising vector field with an L2 objective.
- Flow Matching: The design targets smooth, temporally consistent action chunks rather than highly multimodal distributions, using a small number of Euler steps at inference.
- Inference: The VLM runs once per control step, after which the lightweight action head performs denoising, reducing latency and improving inference throughput.
- Training and Inference Recipe: Training standardizes action and proprioceptive normalization, data shuffling, learning-rate schedules, warm-up, and benchmark-specific action-chunk horizon.
- Training and Inference Recipe: Ablations vary action-transformer scale, VLM backbone, and information-injection mechanisms to contextualize implementation-driven performance differences.
4. Evaluation
SimVLA is evaluated across standard simulation benchmarks and real-robot settings using consistent protocols, with controlled studies of training and implementation choices. It achieves strong benchmark performance, robustness, and real-robot results while showing substantial sensitivity to training details.
- Evaluation setup: SimVLA is evaluated on LIBERO, LIBERO-PRO, and SimplerEnv benchmarks, alongside real-robot experiments, using consistent comparison settings.LIBERO tests long-horizon consistency and generalization; LIBERO-PRO introduces systematic perturbations, while SimplerEnv covers Google Robot and WidowX settings.
- LIBERO: SimVLA achieves a 0.5B-backbone LIBERO average that surpasses OpenVLA-OFT (97.1%) and MemoryVLA (96.7%), with 99.4% Spatial, 99.8% Object, 98.2% Goal, and 96.4% Long success.The model ranks first overall under the matched setup and reaches near-perfect scores on three suites.
- LIBERO-PRO: SimVLA reaches 10.0% success on both Goal and Long task perturbations, while ranking first across LIBERO-PRO Semantic suites at around 98–100% success.Position robustness remains high only on Spatial (29.0%) and low on Object, Goal, and Long.
- SimplerEnv: SimVLA achieves 95.8% overall average success on WidowX, tying X-VLA and surpassing MemoryVLA (71.9%) and FPC-VLA (64.6%) without pre-training.It also obtains 100% on Put Spoon on Towel and Put Eggplant in Basket.
- SimplerEnv: SimVLA achieves 76.1% average success on Google Robot, exceeding SpatialVLA (67.5%), RT-2-X (65.6%), ThinkAct (65.1%), and X-VLA (75.7%).These results are reported for variant aggregation tasks.
- Ablations: Ablations show that data shuffling, action normalization, learning rate, and the VLM learning-rate multiplier can substantially affect LIBERO performance.Changing one training knob while fixing the others can produce drops larger than gains attributed to architectural changes.
- Real-robot evaluation: SimVLA achieves broadly comparable zero-shot real-robot performance to π0.5, with most tasks typically around 80% success.SimVLA is trained end-to-end from a pretrained VLM without VLA or robot-data pre-training, whereas π0.5 uses its publicly released initialization.
5. Conclusion
The paper concludes that SimVLA provides a minimalist, reproducible reference for VLA research. Across simulation and real-world evaluations, it combines strong performance with low memory usage and highlights the importance of standardized implementation details.
- Conclusion: SimVLA addresses performance attribution by decoupling perception from control and using a standardized training recipe.The design is presented as a minimalist baseline for comparing architectural innovations.
- Conclusion: SimVLA rivals or outperforms multi-billion-parameter state-of-the-art baselines across simulation benchmarks and real-world robotic tasks.The evaluations also report scene generalization and a low memory footprint.
- Conclusion: Data shuffling, action normalization, and optimization dynamics can be as influential as architectural novelties.SimVLA is intended to provide a reproducible reference for quantifying the added value of future architectural innovations.
Impact Statement
SimVLA is presented as a transparent and reproducible baseline for VLA research, with potential efficiency benefits and deployment risks. The paper limits its evaluation to controlled research settings and calls for safety constraints, human oversight, and domain-specific validation.
- Scientific impact: Standardized training dynamics and simplified architecture may facilitate fairer comparisons and more reliable progress in embodied AI research.The contribution is framed as a transparent and reproducible baseline.
- Potential benefits: Potential positive impacts include more accessible robotic manipulation systems, efficient assistive-robot deployment, and reduced computational overhead.These impacts are described for domestic or industrial settings.
- Risks and boundaries: SimVLA is evaluated in controlled research settings, so real-world deployment should include safety constraints, human oversight, and domain-specific validation.The paper also identifies risks involving labor automation, unsafe-environment misuse, and deployment without adequate safety validation.
A.1. Galaxea Open-World Dataset
The Galaxea Open-World Dataset is a large-scale real-world mobile-manipulation dataset collected on the Galaxea R1 Lite embodiment. Its aligned multimodal demonstrations support end-to-end VLA training and zero-shot evaluation on the same platform.
- The dataset contains approximately 500 hours of demonstrations, about 100K trajectories, 150 task categories, 50 scenes, more than 1,600 objects, and 58 skills.
- All data are collected with a single consistent embodiment, aligning perception streams, action/state signals, and language annotations for end-to-end VLA training.
- The Galaxea R1 Lite is a 23-DoF mobile bimanual robot with two arms, a torso, an omnidirectional base, and two grippers.
- R1 Lite uses a head stereo RGB camera for scene context and dual Intel RealSense D405 RGB-D wrist cameras for close-range manipulation.
A.2. Real-robot evaluation details
Real-robot evaluation measures binary task completion under a fixed time budget across eight manipulation tasks. The tasks cover object placement, cleaning, folding, disposal, and drawer opening with task-specific success criteria.
- Each real-robot task is scored as binary success or failure over 50 trials within a fixed time budget.
- Store the dolls succeeds when all three plush dolls are fully inside the designated container at episode end.
- Arrange eggs succeeds when one egg is seated in the target carton slot without being dropped.
- Flower, pen, and clothing tasks require stable placement of the flower, containment of both pens, and completion of the target shirt fold on the work surface.
- Wiping, garbage pickup, and drawer opening succeed when the stain is cleaned, both bottles enter the bin, or the drawer passes its opening threshold without damage.
A.3. Training hyperparameters
The training setup summarizes reproducibility-oriented configurations for simulation and real-robot experiments. Simulation uses a pretrained VLM without VLA data pretraining, while the real-robot comparison matches data, hyperparameters, compute, and training steps across methods.
- Simulation: Simulation runs use 4×H100 GPUs and initialize from a pretrained VLM backbone without VLA data pretraining.
- Simulation: The simulation configuration is reported in Table 7, whose caption identifies its hyperparameters for simulation-dataset training.
- Real-robot training: SimVLA and π0.5 are trained on the same Galaxea data with identical hyperparameters, compute budget, and 150K training steps using 64×H100 GPUs.
- Real-robot training: The Galaxea-500h training setup is summarized in Table 8.