Nvidia researchers have introduced HOVER, a small neural network designed to control humanoid robots through several different input methods. The striking part is not only what it controls, but how little model capacity it uses: HOVER has 1.5 million parameters.
That makes the system tiny compared with typical large language models, which use hundreds of billions of parameters. Yet according to the source, HOVER handles complex robot movements and performs better than specialized systems built around one control method.
Why HOVER matters for humanoid robot control
Humanoid robots are difficult to control because movement is not just a matter of sending isolated commands. A robot has to keep balance, coordinate limbs, and respond to a stream of changing input without losing the physical coherence of its motion.
HOVER is presented as a general controller rather than a narrow tool for one device or command style. The same model can take different forms of human input and translate them into robot movement.
The supported inputs listed in the source include:
- Head and hand tracking from XR devices such as Apple Vision Pro
- Full-body positions from motion capture
- Full-body positions from RGB cameras
- Joint angles from exoskeletons
- Standard joystick controls
This range is important because each input source describes human intent differently. A joystick provides a compact command. Motion capture gives body positions. An XR headset may track head and hand movement. Exoskeletons provide joint-angle information. HOVER is designed to work across those formats instead of requiring a separate controller for each one.
Small model, broad behavior
The model size is one of the central facts. HOVER needs only 1.5 million parameters, while typical large language models use hundreds of billions of parameters. The comparison does not mean the systems do the same job, but it gives useful scale: HOVER is compact for the type of physical control it is being asked to perform.
The source says HOVER performs better at each control method than systems designed specifically for just one type of input. That is a notable claim because specialization often brings an advantage in robotics. A controller built for one input stream can be tuned around that exact signal.
Lead author Tairan He speculates that HOVER's broader performance may come from its understanding of physical concepts such as balance and precise limb control. In plain terms, the model may be learning useful movement principles that transfer across input types, rather than only memorizing how to respond to one control format.
That matters because humanoid robot control depends on shared physical constraints. Whether the command comes from a VR headset, a camera, or a joystick, the robot still has to move limbs in a coordinated way and stay balanced.
Training in Nvidia's Isaac simulation
The team trained HOVER in Nvidia's Isaac simulation environment. The source says this environment speeds up robot movements 10,000 times, allowing training to happen far faster than it would in real time.
According to Nvidia researcher Jim Fan, a full year of training in the virtual space takes just 50 minutes of actual computing time on one GPU. That detail highlights why simulation is central to this work. A robot controller can experience a large amount of movement practice in a compressed computing window.
HOVER also moves zero-shot from simulation to physical robots without fine-tuning, according to Fan. In other words, the system can go from the simulated training environment to real robots without an extra adjustment stage described in the source.
That is a meaningful point for robot control because simulation and physical hardware are not the same environment. The source does not provide additional technical detail about how HOVER handles that gap, so the important claim is limited to what is stated: it transfers zero-shot from simulation to physical robots without fine-tuning.
How HOVER fits into existing robotics work
HOVER builds on the open-source H2O & OmniH2O project. It is also described as working with any humanoid robot that can run in the Isaac simulator.
That compatibility framing gives HOVER a broader target than a single machine. The source does not list specific robot models, so the safest way to state the scope is exactly that: the system works with humanoid robots that can run in the Isaac simulator.
Nvidia has also posted examples and code on GitHub. For researchers and developers, that matters because it gives a way to inspect the system and its demonstrations rather than relying only on a high-level description.
The practical takeaway
The main point is simple: Nvidia's HOVER shows that a small neural network can control humanoid robots across multiple input types and reportedly outperform more specialized controllers for those inputs.
The work brings together several important ideas in robotics AI: compact models, simulation-based training, zero-shot transfer to physical robots, and support for multiple human control interfaces. None of those pieces alone solves humanoid robotics, but together they point toward controllers that are less tied to one device or one command style.
For now, the most concrete facts are the ones Nvidia's researchers report: HOVER has 1.5 million parameters, was trained in Nvidia's Isaac simulation environment, can use inputs ranging from XR tracking to joysticks, and builds on the open-source H2O & OmniH2O project.