Frozen 552B VLM: 10 demos beat a 2B policy's 50, charger plugs in 9/10 times

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters

Ruixi Yu, Dairu Liu, Xuchuan Chen, Zekun Qi, Zekai Li, Zhuo Cao, Luyang Niu, Yingsheng Zhang, Sikai Liang, Ke Fan, He Wang, Li Yi

cs.RO

2026-11-17

ScaleVLA freezes a 552B DeepSeek VLM and trains an action DiT. Ten LIBERO-Goal demos hit 87.8%, beating a 2B model on 50 (85.6%); charger insertion goes from 4/10 to 9/10.

What problem this solves

Vision-language-action policies still sit on small backbones. GR00T N1.7 uses the 2B Cosmos-Reason2. GPT-6 Astra can already manipulate objects, and that skill tracks general pretraining rather than a larger pile of robot demonstrations. The controlled question is narrow: hold the action head, the demonstrations, and the optimizer budget fixed, swap only the vision-language backbone, and check whether sample efficiency, out-of-distribution success, and contact precision move together.

Using a frontier model as a planner means a text pass and an action review on every decision, which is slow. A small backbone often reaches the socket and still misses the insertion. ScaleVLA freezes the VLM, reads it once per action chunk, and never generates text in the control loop.

Method

The backbone is DeepSeek-V4.1-Flash-1017, a 552B mixture-of-experts model. Vision and language weights stay frozen. Images and the instruction go through the model's own chat template into a 20-layer causal encoder. ScaleVLA reads the tokens that would enter the next attention layer, each of width 5120. The 20-layer decoder never runs.

The action head follows GR00T N1.7 and is trained from scratch. Layer norm plus a two-layer projection maps backbone tokens to 2048 dimensions, then a 4-layer self-attention block mixes them under the action loss. Embodiment-specific MLPs encode state and the noisy action. The action DiT, a diffusion transformer that denoises a block of actions, has 32 layers: self-attention over state and action tokens, and cross-attention alternating between non-image tokens and image tokens. Flow matching draws a straight path from noise to the demonstration and regresses its velocity, with squared error only on valid action dimensions. Frozen features are precomputed once.

At inference the backbone runs once per chunk, a short block of future actions. The DiT integrates the velocity in four steps on those same features. LIBERO predicts 16 steps and executes 8 before replanning. DexJoCo predicts 30 steps and generates the next chunk while the controller is still executing. The comparison also includes dense Qwen3-VL (2B, 32B) and Cosmos-Reason2 (2B, 8B, 32B). Action-head architecture, random initialization, and training budget stay the same. No pretrained robot action expert is loaded.

Results

The sample-efficiency sweep fixes training at 5,000 steps on nested demonstration subsets. The 2B, 8B, and 32B rows are Cosmos-Reason2.

Demos per taskGoal 2BGoal 552BLong 2BLong 552B
539.4%77.4%24.0%62.2%
1055.0%87.8%56.0%75.0%
2585.2%97.6%69.6%84.6%
5085.6%95.4%82.4%86.8%

On LIBERO-Goal, 10 demos with the 552B backbone reach 87.8%, above the 2B backbone trained on 50 demos (85.6%). Five demos move Goal from 39.4% to 77.4%. At that same budget, 8B scores 60.6% and 32B scores 61.2%, so the mid-scale models barely separate. At 50 demos the 552B Goal score (95.4%) ties the 32B score (95.2%) and sits below its own 25-demo score of 97.6%. On Long, 50 demos give the 8B model 87.2%, a hair above the 552B model at 86.8%.

With the full LIBERO demonstration sets, average success is 96.1% for ScaleVLA, 94.0% for Cosmos-Reason2-32B, 93.6% for Qwen3-VL-32B, 93.7% for Cosmos-Reason2-8B, 91.2% for Qwen3-VL-2B, and 90.7% for Cosmos-Reason2-2B. Published averages are 94.2% for π0, 76.5% for OpenVLA, and 83.9% for CoT-VLA. Spatial and Object are near saturation. The 552B gains sit on Goal (96.8%) and Long (90.2%).

LIBERO-PRO applies environment and object shifts with no retraining. ScaleVLA averages 45.1%. The best dense backbone, Cosmos-Reason2-8B, averages 31.3%. The two 2B models average 29.3% and 23.3%. Qwen3-VL-32B averages only 22.8%, below the 8B model. On Goal, ScaleVLA scores 47.8% under environment shifts and 56.4% under object shifts. On Object environment shifts it scores 31.2%, below Cosmos-Reason2-32B at 32.4%.

On RoboCasa's out-of-distribution kitchens, 30 generated demos per task put the 2B to 32B models at 22.4% to 25.6% and ScaleVLA at 30.7%. With 100 demos the smaller models sit at 42.8% to 45.6% and ScaleVLA at 54.9%.

DexJoCo trains one policy on five bimanual tasks. Under object randomization, average success is 31.9% to 37.1% at 2B, 38.1% at 8B and at 32B, and 42.1% for ScaleVLA. Full randomization drops ScaleVLA to 24.4%, against 22.1% for the best smaller backbone. Assembly stays near failure, at most 1.3%. On Microwave with object randomization, ScaleVLA scores 73.3%, under Qwen3-VL-32B at 89.3%. The published DP-T and π0.5 averages of 16.0% and 10.1% come from a different training setup.

The real task is bimanual charger insertion. The right arm grasps and hands the plug over, and the left arm aligns it with a power strip and inserts it. Cameras are a top view plus two wrists. Most demonstrations start near one reference pose. The π0.5 backbone has seen large-scale robot data. Validation action error is similar for the 2B and 552B policies.

PolicyIn-domainOODInsertionCompletionLatency
Cosmos-2B4/105/1585.9 s147.1 s311 ms
π0.58/107/1581.2 s132.0 s262 ms
ScaleVLA9/108/1558.1 s114.4 s384 ms

In-domain failures all happen at insertion. ScaleVLA's median insertion is 58.1 s, 28% shorter than π0.5 at 81.2 s. Median completion is 114.4 s, 13% shorter than 132.0 s. The 2B policy hits repeated safety stops from insufficient arm clearance. ScaleVLA does not.

On RoboCasa, ScaleVLA has 276 times the parameters of Cosmos-Reason2-2B and about 1.9 times the median chunk latency, 216 ms versus 115 ms. The action head costs about 70 ms, so growing a dense backbone from 2B to 32B adds only about 26 ms. ScaleVLA uses four H200 GPUs; the others use one. It also consumes 568 tokens per observation against 212, and that share of the latency is not isolated. Round-trip latency on the robot is 384 ms.

Why it matters

Training cost sits in the action head. Once the backbone is frozen, features can be cached, so a backbone ablation does not mean training 552B parameters. The payoff shows up when demonstrations are scarce, the scene shifts, or the plug has to seat. Saturated pick-and-place on LIBERO is already served by an 8B or 32B backbone.

π0.5 asks whether robot-data pretraining can stand in for scale. It reaches 8/10 in-domain and 7/15 out of domain, which closes much of the success-rate gap, and it still takes about 23 seconds longer to insert. Robot pretraining covers part of the task. Contact alignment is where the frontier representation still leads.

Sub-second chunks are practical on four H200s, because the backbone runs once per chunk and not once per servo tick. The higher-latency 552B policy finishes the task sooner. The saved time is fewer failed alignments.

Limitations

The real-robot evidence is one task, ten in-domain trials and fifteen shifted ones. Out of domain, success is 8/15, and the demonstrations cluster around a single reference pose. A 9/10 in-domain rate is not an open-table insertion result.

The four-suite LIBERO average moves from 94.0% on the best 32B dense backbone to 96.1% at 552B, and Spatial and Object are already saturated. The readable gaps are Goal, Long, and the perturbation suite. On Object environment shifts, the 552B model does not beat the 32B model. DexJoCo under full randomization is a 2.3-point average edge, and Assembly barely works. Microwave is a case where 552B is worse. At 50 Goal demonstrations the 552B model ties 32B and falls short of its own 25-demo score of 97.6%. The low-data story does not extend to the data-rich setting.

Latency also mixes in token count, 568 tokens per frame against 212. The paper only says a fraction of the experts fire, without stating the active parameter count. 552B is the total size. Per-step compute is not reported.

There is no head-to-head with Astra-style test-time reasoning. The backbone never receives a gradient. Whether unfreezing the top layers, or continuing pretraining on robot data, would move the numbers further is unmeasured.

Terms

Source

Related papers

All paper explainers