TurboVLA Runs Robot Control at 32 Hz Without a Giant LLM

Many vision-language-action systems route camera features and instructions through a multi-billion-parameter language model before generating robot actions. That design brings useful semantic knowledge, but the language backbone remains active every time the policy is queried, adding memory use and latency.

TurboVLA uses a more direct route. DINOv3 encodes the camera images, BERT encodes the instruction, and six bidirectional interaction layers exchange information between the two streams. An ACT-style decoder then predicts a chunk of continuous robot actions in one forward pass. The model still uses language; it avoids a large generative language model at the center of low-level execution.

The paper reports 31.2 milliseconds from receiving the multimodal observation to producing an action chunk on an NVIDIA RTX 4090 with batch size one. That corresponds to roughly 32 policy predictions per second. The comparison uses the authors’ implementations and official checkpoints for runnable baselines on the same GPU.

This is policy-inference frequency, not proof that every part of the robot operates at 32 Hz. Camera capture, network transport, safety checks, command execution and the arm’s lower-level servo controller add their own timing. The result still matters because a slow policy can leave the robot acting on stale visual information.

TurboVLA was evaluated on the four LIBERO suites covering object, spatial, goal-conditioned and longer-horizon manipulation. The study used 50 rollouts for each of 40 tasks, producing 2,000 trials in total. The reported average success rate was 97.7%.

The 216.1-million-parameter configuration used 0.9 GB of inference VRAM and 31.2 milliseconds per policy call. In the paper’s same-hardware comparison, π0.5 reached 96.9% with 3.4 billion parameters, 12.8 GB of VRAM and 93.6 milliseconds of latency. These are benchmark results under the paper’s protocol, not a universal ranking across every robot and environment.

The physical setup used an AgileX Piper arm with a wrist-view RGB-D camera and a third-view RGB-D camera. The four instructions were to grab a roller, move a playing card away, press a stapler and stack three bowls. The model was initialized from LIBERO pretraining and fine-tuned with 65 teleoperated demonstrations for each task.

Each task was evaluated over 40 trials. The paper reports success rates of 92.5% for grabbing the roller, 80% for moving the playing card, 90% for pressing the stapler and 87.5% for stacking three bowls. TurboVLA outperformed π0.5 under the same platform, data and evaluation protocol according to the authors.

For RoboTwin 2.0, the team trained one policy across 50 clean bimanual tasks. This version uses a DINOv3 ViT-L visual backbone, three camera views, 14-dimensional joint-position actions and a 50-step action chunk.

The bimanual configuration is reported at 0.4 billion parameters rather than 0.2 billion. It achieved 60.2% average success with 43.4 milliseconds of latency, compared with 57.0% for π0.5 and 50.3% for StarVLA-α in the paper’s clean-setting comparison. Randomized-scene data were excluded because of the authors’ compute budget.

TurboVLA is strong evidence that bounded execution does not always need a giant language model. A command such as “press the stapler” already defines the task. The controller mainly needs to locate the object, estimate the scene and generate a stable movement.

High-level planning is different. A household robot asked to clean an unfamiliar room may need to interpret an ambiguous request, decompose it into steps, choose tools and recover from unexpected conditions. The authors explicitly state that TurboVLA is designed for concrete execution-level instructions and may lack the semantic reasoning required for high-level task planning. This is why parameter count alone does not decide whether a system qualifies as a broad robot foundation model.

Removing language reduced the reported LIBERO average from 97.7% to 70.8%. The largest collapse occurred in goal-conditioned tasks, where success fell from 97.4% to 11.6%. A learned task-ID embedding recovered much of the performance but remained below natural-language conditioning.

The text encoder did not need to be a giant generative model. T5-Small reached 97.1%, while the chosen BERT setup reached 97.7%. Bidirectional interaction between vision and language also performed better than simple concatenation or one-way attention in the reported ablations.

The official repository publishes training and evaluation code under Apache-2.0. The team also released LIBERO and RoboTwin checkpoints with normalization files and checksums. The checkpoints include DINOv3-derived parameters and are distributed under additional DINOv3 terms, so the code license does not describe every released weight.

The evidence remains a first-version preprint and author-run evaluation. The real-world tasks are controlled tabletop experiments on one arm, with 160 trials in total. The paper does not establish long-duration reliability, unseen-home performance, safety around people, operation on embedded robot compute or independent reproduction of the physical trials.