wtf is hoic-baseline?

This repository provides two baseline methods for testing how humanoid robots can interact with physical objects: pushing and pulling carts, moving boxes across stairs and slopes, riding a skateboard, and maneuvering a wheelchair. It accompanies research papers from 2026 and is designed so that other researchers can run these simulations and measure success rates on the provided tasks. The two methods are called HAIC and VAIC. HAIC works by giving the robot reference motion clips (recordings of how a robot-object pair should move) along with a model that tracks the physics of the interaction. The robot uses those references to figure out how to control itself. VAIC takes a different approach: it is reference-free and instead uses a depth image from the robot's own perspective to adapt to objects it has not seen before. The simulation runs inside MuJoCo, a widely used physics engine for robotics research. Setup requires Python 3.8 with a few packages (mujoco, onnxruntime, opencv, and a couple of image tools). The trained policies are stored as ONNX files, which are a portable format for sharing neural network models, so you do not need to retrain anything to run the demos. You simply activate a conda environment, pick a task from the list (skateboard, box_plane, push_cart, push_wheelchair, and several others), and run the corresponding script. The repository also includes options for recording videos of the simulation runs and for adjusting timing parameters such as physics timestep and control frequency. A helper script checks that every task has its policy file, motion file, and robot description in place before you try to run it. This is a research baseline, meaning it is meant as a starting point for building on top of these methods, not a finished product. The VAIC policy deployment is noted as not yet released. If you want to train your own policy using the same framework, the authors link to separate repositories for HAIC and VAIC training code.