TL;DR
ScaleVLA demonstrates that scaling the Vision-Language Model (VLM) backbone of a robot policy to 552 billion parameters significantly boosts sample efficiency, generalization, and manipulation precision. By connecting a massive DeepSeek-V4.1-Flash backbone to a trainable action head, the researchers achieved state-of-the-art results on benchmarks like LIBERO-PRO and real-world bimanual tasks, all while maintaining a practical 216ms inference latency.
What Problem Does This Paper Address?
In the quest for "foundation models" for robotics, most current Vision-Language-Action (VLA) policies rely on relatively small backbones (typically 2B to 7B parameters). While these models are effective, they often struggle with out-of-distribution (OOD) generalization and require large amounts of robot-specific data to learn new tasks.
Meanwhile, frontier VLMs like GPT-6 Astra have shown incredible "world knowledge," but they are often used as high-level reasoning agents that generate text-based plans—a process that is too slow and computationally expensive for real-time, low-level motor control. The authors of ScaleVLA ask a fundamental question: Can we use frontier-scale representations as a direct foundation for robot actions? Specifically, does scaling the backbone to hundreds of billions of parameters actually translate to better physical performance, or is there a point of diminishing returns?
Methods
The researchers introduced ScaleVLA, a framework that bridges the gap between massive language models and continuous robotic control.
- The Backbone: They utilized DeepSeek-V4.1-Flash, a 552B parameter model using a sparse Mixture-of-Experts (MoE) architecture. This choice is critical; MoE allows the model to have a massive parameter count while only activating a fraction of them during any given inference step, keeping latency manageable.
- The Action Head: Instead of making the model "reason" in text, they connected the VLM's internal representations to a Diffusion Transformer (DiT). This head is trained from scratch to translate multimodal features into continuous robot actions using flow matching.
- Frozen vs. Trainable: To isolate the effect of the backbone, the VLM remains frozen. Only the "feature adapter" (which maps VLM tokens to the action head's space) and the action head itself are trained on robot demonstrations.
- Non-Autoregressive Control: ScaleVLA reads the backbone once per action chunk rather than generating actions token-by-token. This architectural choice is what allows a 552B model to run at sub-second speeds.
Datasets
The model was evaluated across a rigorous suite of simulation and real-world benchmarks:
- LIBERO & LIBERO-PRO: Standard tabletop manipulation tasks, with the "PRO" version introducing environment and object perturbations to test robustness.
- RoboCasa: Kitchen tasks set in out-of-distribution scenes.
- DexJoCo: Five complex bimanual dexterous tasks (e.g., unlocking an iPad, assembly).
- Real-World Bimanual: A "Plug In Charger" task requiring a handover between arms and precise, contact-sensitive insertion into a power strip.
Key Insights and Ideas
- Scaling as a Data Multiplier: The most striking finding is the impact on sample efficiency. ScaleVLA with only 10 demonstrations outperformed a 2B parameter backbone trained with 50 demonstrations. This suggests that larger models possess a "common sense" about physics and space that reduces the need for repetitive robot-specific training.
- Generalization through Scale: On LIBERO-PRO, where environments are intentionally changed from the training set, ScaleVLA maintained a 45.1% success rate, while smaller backbones plummeted to ~31%. The "world knowledge" from large-scale pretraining acts as a buffer against environment shifts.
- Precision is a Representational Problem: Gains weren't just in high-level logic. In real-world charger insertion, ScaleVLA excelled at the final, contact-rich alignment phase. This indicates that frontier VLMs provide higher-fidelity spatial representations than their smaller counterparts.
- The Latency Paradox: You might expect a 276x increase in parameters to lead to a 276x increase in latency. However, due to the MoE architecture and the non-autoregressive action head, ScaleVLA's latency was only ~1.9x higher than a 2B model (216ms vs 115ms).
Key Results
- LIBERO-PRO: Achieved 45.1% average success, a significant jump over the 31.3% achieved by the best dense 32B backbone.
- Sample Efficiency: On LIBERO-Goal, ScaleVLA reached 87.8% success with 10 demos, surpassing the 2B baseline's 85.6% with 50 demos.
- Real-World Performance: In the bimanual charger task, ScaleVLA succeeded in 9/10 trials, whereas the 2B backbone only managed 4/10. ScaleVLA was notably faster at the insertion stage, completing trials 13% faster overall because it didn't struggle with alignment.
- Inference Speed: Achieved a median latency of 216 ms per action chunk using four H200 GPUs.
Limitations and Future Work
- Hardware Requirements: While latency is low, the "cost of entry" is high. ScaleVLA requires four H200 GPUs to run effectively, whereas smaller models can run on a single, less powerful card.
- Long-Horizon Reasoning: While the model is excellent at immediate action, the authors note that for extremely long-horizon tasks, additional task-specific supervision or higher-level planning might still be necessary.
- Future Directions: The authors suggest exploring different MoE configurations and investigating if even larger scales (trillion-parameter models) continue to yield linear improvements in robotic precision.
Summary updated