We introduce three core techniques to our RL-based training framework to significantly expand the operational workspce of humanoid robots whereas ensuring robustness of locomotion:

1. Upper-body pose curriculum: Enable balance under continuous changing upper-body poses.
2. Height tracking reward: Enable the robot to squat to any required heights robustly and quickly.
3. Symmetry utilization: Make the robot act more symmetrically & improve data efficiency.

Our framework is totally MoCap-free, resulting in a more efficient pipeline.

Our framework can be used to train different kinds of robots such as Unitree G1and Fourier GR-1.

After training with our framework on an Nvidia RTX 4090 for only about 3 hours , we can get policies that can be deployed directly in the real world to drive robots walk and squat robustly.

We conduct serval ablation experiments to verify the effectiveness of our framework, and we find:

1. Our upper-body pose curriculum can help robots better learn to balance under dynamic upper-body movements gradually than methods without curriculum or with other curriculum style.
2. The introduction of novel height tracking reward can accelerate the training for robot squatting.
3. The symmetry utilization can both significantly accelerate the training process by over 10 times and guarantee the symmetry of the trained policy.

Our hardware system features isomorphic exoskeleton arms, a pair of motion-sensing gloves, and a pedal. The pedal design for locomotion command acquisition liberates the operator's upper body, enabling simultaneous acquisition of upper-body poses. Since the exoskeleton arms are isomorphic to the controlled robot and each glove has 15 degrees of freedom (DoF), which is more than most existing dexterous hands, we can directly set upper-body joint positions from the exoskeleton readings, dispensing with IK and achieving faster and more accurate teleoperation.

We design hardware systems for both Unitree G1 and Fourier GR-1. Notably, our gloves can be detached from the arms, allowing them to be reused in systems isomorphic to different robots.

Using our hardware system, one single operator can choose to control:

We deploy the trained policy on the Unitree G1 in the real world and teleoperate it to perform various loco-manipulation tasks using our isomorphic exoskeleton hardware system.

We further conduct some experiments to show the robustness of our policy.

To demonstrate the effectiveness of the isomorphic exoskeleton, we compare the task completion times across four different tasks between our hardware system and OpenTelevision.

The completion times for these tasks are computed based on data from three different operators, with each operator performing the tasks three times. Our hardware system can accelerate the teleoperation by approximately 2 times, particularly in tasks that require radial movement.

1. Simulation

We transfer the trained policies for Unitree G1 and Fourier GR-1 from Isaac Gym to scenes developed by GRUtopia, thus the robots can perform diverse loco-manipulation tasks more cost-effectively and in a wider range of scenarios than would be feasible in the real world.

2. Imitation Learning

To validate the effectiveness of the demonstratons collected by HOMIE for IL algorithms, we design two distinct tasks, collect data by teleoperating, train with IL algorithm, and deploy in the real world. We achieve over 70% success rate, showing the feasibility of training IL with collected data.