TurboVLA, NeoRacer and HumanCLAW | AI robotics

What actually changed in AI robotics in the last 24 hours? Three useful things showed up, but none is a finished household robot. There is a smaller VLA with released code, an open racing car that is still taking pre-orders, and a benchmark that puts a number on a hard problem: choosing the next action with a physical body.

TurboVLA: a smaller path from instruction to action

TurboVLA skips the usual large language model in the middle of a vision-language-action policy. It uses separate vision and language encoders, then predicts continuous action chunks with a compact decoder. The authors also released the training and evaluation code in the official repository.

Their paper reports a 0.2B-parameter model reaching 97.7% average success on LIBERO, with 31.2 ms inference latency and 0.9 GB of VRAM on an RTX 4090. Those are author results on a benchmark, not an independent test on a home robot. Still, the useful bit is concrete: a lower inference budget makes it more plausible to spend a consumer GPU budget on data collection and iteration instead of only on running the policy.

NeoRacer: open hardware, not a delivered product yet

NeoRacer is a documented 1:12 autonomous racing platform aimed at education and repeatable research. Its paper specifies a Jetson Orin Nano, 270° LiDAR, a 120 fps global-shutter camera, an IMU and encoders. The authors say the hardware is licensed under CERN-OHL-S v2 and the software under GPLv3.

The Neobotics pre-order page lists an early price of $2,500, then $2,700, and says first-batch shipping is expected in June or July. That is a company statement, and I found no independent confirmation that units have shipped. Treat it as a pre-order, not a field deployment.

For a builder, the better lesson is its sequence. Start with manual RC control, add a camera, then perception, then autonomy. The paper’s shared hardware layout is intended to make comparisons more repeatable.

HumanCLAW: planning is still the weak link

HumanCLAW is a benchmark designed to separate a vision-language model’s action choice from low-level motor errors. It contains 1,218 egocentric episodes across 41 indoor scenes, and tests nine models on finding, navigating and interacting.

The best model reaches 16.8% success in the paper. It is a preprint, so it has not had peer review or an outside replication. The result is still useful because it narrows the failure: the models struggle to keep track of their own body, goal and obstacles, even when the control layer is held apart from the decision.

If you are building with AI, keep the boring base in place. Sensors, odometry and safe control decide whether the machine knows where it is. Let the model help choose an action after that. Start with one reliable behaviour and build from there.