This paper aims to pioneer the deployment of Proximal Policy Optimization (PPO) on resource-constrained RISC-V platforms to address the critical dependency of industrial humanoid gait control on expensive high-performance processors. The self-developed 16-DoF robot (dual-core 1 GHz C906/512MB RAM) overcomes real-time control limitations in unstructured environments like warehouses and disaster sites.
Through hardware–software codesign: Initially, an RISC-V hardware platform is constructed using 3D-printed frames and ST3215 servo networks. Subsequently, PPO model size is reduced by 91% (from 10MB to 0.9MB) via mixture Gaussian outputs and curriculum learning. Finally, a Diffusion Policy is innovatively integrated, where spatiotemporal U-Net resolves joint coordination dependencies and two-stage sampling optimizes high-dimensional action spaces.
Experiments show that diffusion-based PPO achieves 87 frames per second real-time control on the RISC-V platform (11.5 ms latency), reducing lateral offset by 89% over a 3-m walk (1.5 cm vs 13.9 cm) and trunk tilt by 75% compared to baseline PPO. In addition, it reduces hardware costs by $200 per unit compared to ARM-based solutions. Tests on complex terrain further demonstrate superior robustness, with maximum lateral offset of 4.8 cm and average trunk tilt of 4.3° on inclined slopes with disturbances.
Breakthrough contributions include, to the best of the authors’ knowledge, the first real-time deployment of a diffusion-based PPO for humanoid gait control on a low-cost RISC-V platform; a novel Diffusion–RL hybrid for high-dimensional joint coordination; and a 5.1× convergence boost via biomechanical constraints. This offers a cost-effective and reproducible paradigm for resource-constrained edge robotics.
