ZICQ
中 Log in / Sign up
Newsroom Research & Papers #ScaleVLA #Vision-Language-Action #Robot Learning #Action DiT #DeepSeek-V4.1-Flash

ScaleVLA Scales Vision-Language-Action Models to 552B Parameters

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 1 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:ScaleVLA connects the paper's 552B DeepSeek-V4.1-Flash vision-language backbone to a trainable Action Diffusion Transformer and compares backbones ranging from 2B to 552B parameters under a shared training setup. The reported results show 87.8% success with only 10 demonstrations, 45.1% success on LIBERO-PRO, and an improvement from 4/10 to 9/10 on real-world charger insertion, while maintaining a 216 ms median latency per action chunk. The work provides evidence that frontier-scale visual-lang


ScaleVLA tests a more specific question than whether a larger model can improve a robot's headline success rate: can representations learned by a frontier-scale vision-language model provide a stronger foundation for action policies, rather than serving only as reasoning or planning agents?

Architecture: separating representation learning from action generation

The system connects a 552B-parameter VLM, identified in the paper as DeepSeek-V4.1-Flash, to a trainable Action Diffusion Transformer. The VLM encodes visual observations and task language into semantic representations, while the Action DiT uses those representations to generate continuous action chunks. A conventional DiT implementation denoises a noisy action sequence over multiple sampling steps, although the supplied material does not disclose the action dimensionality, chunk horizon, DiT depth, diffusion schedule, normalization scheme, or whether the 552B backbone is frozen.

This separation places the expensive computation in general-purpose representation extraction and leaves continuous control generation to a dedicated action head. Whether that actually reduces total cost depends on execution details. If the VLM runs only once per observation chunk, its forward pass can be amortized across several actions. If the backbone is repeatedly invoked during diffusion sampling, the latency profile changes substantially. The reported 216 ms median per action chunk does not provide enough information to reconstruct the complete inference graph.

The study also compares VLMs ranging from 2B to 552B parameters under a shared training setup. Holding the training recipe relatively constant makes the scaling trend easier to interpret, but it does not fully separate parameter count from differences in pretraining corpus, architecture, active parameters, visual encoder, tokenizer, and action-head capacity.

Reported results

| Capability | ScaleVLA result | Comparison | Interpretation | |---|---:|---:|---| | Demonstration efficiency | 87.8% with 10 demonstrations | 85.6% with 50 demonstrations using a 2B backbone | Five times less demonstration data, but only a 2.2 percentage-point gain | | Out-of-distribution generalization | 45.1% success on LIBERO-PRO | At most 31.3% for smaller backbones | A 13.8-point absolute gain, with substantial remaining failure rate | | Real-world precision | 9/10 charger insertions | 4/10 | Improvement concentrated in contact-sensitive alignment | | Online execution | 216 ms median latency per action chunk | Not reported | Approximately 4.63 chunks per second, not the same as individual control frequency |

The sample-efficiency result is the paper's most compelling signal. Large-scale visual-language pretraining may allow a policy to extract useful priors from substantially fewer robot demonstrations. It should not, however, be interpreted as a fivefold reduction in training cost. A 552B model still incurs far more pretraining and inference computation than a 2B model, even if fewer task-specific demonstrations are required.

The 45.1% result on LIBERO-PRO confirms that scale improves generalization but does not solve it. The 13.8-point absolute gain is substantial, yet a success rate below 50% remains inadequate for many contact-rich or safety-sensitive deployments.

The charger-insertion result also needs statistical caution. With only ten trials, the uncertainty intervals are wide. A rough Wilson-interval calculation gives approximately 17%—69% for the 4/10 baseline and 60%—98% for the 9/10 result. More trials, varied charger instances, and multiple random seeds are required before treating the improvement as stable.

Engineering trade-offs

If the 552B backbone is dense, weight residency alone becomes a deployment barrier. Using memory ≈ parameter count × bytes per parameter, BF16 weights require roughly 1.10 TB, FP8 requires about 552 GB, and INT4 has a theoretical lower bound near 276 GB. These figures exclude quantization metadata, the visual encoder, Action DiT, KV cache, runtime workspaces, and communication buffers. If the backbone is MoE, activated parameters determine much of the compute cost, but the complete expert set may still need to remain resident or be distributed across nodes, adding routing and All-to-All overhead.

Training cost depends on whether the VLM participates in backpropagation. The description explicitly identifies the Action DiT as trainable but does not clarify the status of the backbone. A frozen VLM would reduce gradient and optimizer-state costs, although all weights would still need to be loaded for inference. Full fine-tuning under mixed-precision Adam can require roughly 12 bytes per parameter for parameter, gradient, and optimizer state alone, or about 6.6 TB for 552B parameters, before activations and communication costs are included.

The 216 ms figure is a median. Robot applications also need p95 and p99 latency, jitter measurements, and explicit failure-recovery behavior. Action chunking can amortize VLM calls, but the effective control rate depends on chunk length, observation refresh frequency, and whether the controller replans in a rolling horizon. Diffusion sampling steps, visual preprocessing, network transport, and safety-controller overhead are not described.

Semantic pretraining also does not automatically provide contact dynamics. The material does not say whether the system includes force torque, tactile, depth, or proprioceptive observations. For insertion, assembly, and deformable-object tasks, VLM priors can improve object and task recognition, but the action head must still model friction, contact forces, actuator error, and compliant interaction.

Evaluation quality and remaining confounds

A shared training setup is a meaningful strength, but a strict causal claim about scaling requires additional controls:

  • Fix pretraining data volume, visual encoder, action-head capacity, and training-token count while varying model size.
  • Report equal-data, equal-training-compute, and equal-inference-compute comparisons separately.
  • Publish multi-seed results, episode counts, confidence intervals, and failure-mode distributions.
  • Add parameter-matched generic VLMs, non-VLM visual backbones, and language-free ablations to determine whether the gains come from language semantics or raw capacity.
  • Document the composition of LIBERO-PRO, check for contamination, and test sensitivity to initialization.
  • Include visual preprocessing, action decoding, communication, and safety layers in end-to-end latency measurements.

Without test-set size and variance, the 2.2-point difference between 87.8% and 85.6% cannot be judged for statistical significance. The 13.8-point LIBERO-PRO gap is more consequential and warrants independent reproduction.

Practical deployment guidance

  1. Reproduce the scaling curve before attempting 552B-scale deployment. Test 2B, 7B, 14B, and 32B backbones from the same family while fixing action data, training recipes, and evaluation protocols.
  2. Separate data efficiency from compute efficiency. Track pretraining cost, policy-training tokens, demonstration-collection cost, inference energy, and operational latency rather than relying on demonstration count alone.
  3. Start with frozen backbones, LoRA, or adapters. If full backbone updating is not the source of the gain, training only the action head or lightweight adapters can reduce storage requirements and catastrophic forgetting.
  4. Use chunked actions with receding-horizon control. Refresh observations continuously and enforce collision, force, and velocity safety envelopes; median latency alone is not an adequate control guarantee.
  5. Add sensing and data targeted at contact. Insertion tasks should explicitly incorporate tactile, force-torque, or depth observations and prioritize contact-phase alignment failures and recovery trajectories.
  6. Require production-grade engineering disclosures. Report weight-loading strategy, hardware topology, batch-size-one p50/p95/p99 latency, memory use, power, throughput, fallback behavior, and reproducible configurations.

Bottom line

The meaningful contribution of ScaleVLA is not the claim that every robot needs a 552B model. It is evidence that stronger visual-language pretraining can improve demonstration efficiency, generalization, and contact-sensitive manipulation under a shared training setup. The deployment cost of such scale remains substantial, making smaller VLMs with parameter-efficient adaptation, action-head distillation, and specialized continuous-control modules the more practical near-term path. The current results are an important research signal, not yet a fully validated robot foundation model.


Source: Hugging Face Daily Papers (2026-11-16)

— END —

Tags: #ScaleVLA #Vision-Language-Action #Robot Learning #Action DiT #DeepSeek-V4.1-Flash

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…