TurboVLA Hits 32Hz With 0.9GB of VRAM

Featured

01 Robotics | Real-Time VLA Fits Under 1GB

Running at 32Hz on an RTX 4090 with just 0.9GB of inference memory changes the hardware assumptions for high-frequency robot control. Server-class GPUs may no longer be the default requirement.

TurboVLA does not further compress a large language model. It bypasses the conventional “vision → language model → action” pipeline. Vision and instructions are encoded separately, then fused through lightweight bidirectional interaction. A compact decoder directly predicts continuous action chunks.

The result is a 0.2B-parameter model with 31.2-millisecond latency. It averages 97.7% success on LIBERO, matching or beating larger VLA policies. This reopens the architecture question. Robots that only need to understand tasks and act promptly may not need every decision routed through an LLM representation space.

The design could bring faster closed-loop control to workstations and edge devices. The abstract alone cannot show how it handles complex instructions, unfamiliar scenes, or cross-task generalization. Those questions need the full paper and more physical-robot testing.

Key takeaways: - Real-time VLA can come from redesigning the compute path, not only compressing the language model. - At 0.9GB of VRAM and 31.2-millisecond latency, consumer-grade deployment looks increasingly feasible. - Test lightweight architectures carefully on complex instructions and out-of-distribution scenes.

Source: TurboVLA: Real-Time Vision-Language-Action Model at 32Hz on an RTX4090 with <1GB VRAM

02 Evaluation | Was It the Brain or Body?

Embodied-task benchmarks often misclassify failures. A model may choose the right route, yet the robot falls after losing balance. The final score still records a decision failure.

HumanCLAW separates these causes. The vision-language model issues one atomic skill command at a time. The system converts each command into less than one second of whole-body motion. It also minimizes errors from balance, disturbances, and motors.

This setup isolates action judgment. Does the model know where its body is? Can it tell whether it reached the target or hit an obstacle? None of nine advanced models passed the 1,218 long-horizon tasks. The best success rate reached only 16.8%. Target recognition was not the main bottleneck.

Embodied-AI teams can use this framework for failure attribution. First determine whether the system thought incorrectly or executed poorly. Then decide whether to invest in model training or motion control.

Key takeaways: - Separate high-level decisions from low-level execution to avoid optimizing the wrong component. - Body-state tracking appears weaker than target recognition in current models. - Audit arrival detection, collision sensing, and self-localization first when debugging navigation and interaction failures.

Source: HumanCLAW: Can Vision-Language Models Act Through a Body?

03 Agent | Cheap Tasks Can Still Fail

Combining labor time, task pricing, and inference cost changes the deployment question. Completion rate alone is not enough. The real test is whether output quality justifies the cost.

OmegaUse-OfficeVal contains 100 long-horizon office tasks drawn from practitioner requests. Humans need 2.32 hours on average to complete each one. Every task includes labor-time and price proxies for direct comparison with model inference costs.

The benchmark also builds code-based validators from detailed scoring rules. This aims to make evaluations of complex deliverables more consistent. Frontier models finish faster and cost less than humans, but their output quality still falls well short.

Low inference cost does not yet translate directly into strong automation returns. With only 100 tasks, the benchmark offers a new evaluation framework rather than a final verdict on office automation ROI.

Key takeaways: - Evaluate office agents across quality, labor time, and inference cost in the same accounting model. - Faster and cheaper execution does not mean an agent can replace human delivery. - Use economic evaluation to select deployment targets, but do not generalize ROI from this sample alone.

Source: OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding