HC2Robot: Advanced Humanoid Robotics for Complex Terrain Navigation

Paper: HC2Robot: Hyper-Connection Transformers with Predicted Foresight for Multi-Terrain Humanoid Parkour

Authors: Wenhao Xu, Yanheng Mai, Yifei Fu, Nan Zheng, Yanbo Zhou, Zirui Huang, Mingsheng Ling, Xinjue Wang, Renjing Xu

Executive Summary

Recent advancements in humanoid robotics have focused on enhancing the ability of robots to navigate complex terrains, such as those encountered in parkour. This research addresses a specific challenge in robot locomotion: the disparity between the information available during training and what is accessible in real-world scenarios. In traditional methods, a learning model (the critic) uses detailed data from simulations to optimize performance, while the robot (the actor) must operate with limited, onboard sensory inputs once deployed.

The study introduces a new framework called HC2Robot, designed to bridge this gap. Instead of simply sharing privileged information from the simulator with the actor, HC2Robot creates a deployable approximation of that information. This approach enables the robot to make informed decisions about its movements without relying on real-time, extensive sensory data.

At the core of HC2Robot is a Predicted Foresight Module (PFM). This module predicts the timing, location, and quality of each foot's next landing based on the robot's current observations. By transforming complex simulator data into a simplified, usable format, the PFM equips the robot with crucial information about potential footholds, facilitating safer and more effective movement.

The HC2Robot framework also incorporates a Hyper-Connection (HC) Transformer, which improves the robot’s ability to reason about time and movement. It uses multi-stream routing to process various input signals, further refining the robot's decision-making capabilities. Alongside this, a future motion prediction component provides short-term movement guidance, enhancing the robot's performance even in challenging terrains.

In simulated environments that include stairs, gaps, high platforms, and slopes, HC2Robot demonstrated a remarkable success rate of 97.60% in traversing these obstacles. It also consistently outperformed existing locomotion systems in terms of identifying safe footholds. These findings suggest that HC2Robot could significantly improve the efficiency and safety of humanoid robots operating in unpredictable environments.

This research is particularly relevant for industries that deploy robotic systems in dynamic settings, such as search and rescue operations, entertainment, or any field requiring agile, humanoid robots. While the results stem from simulations rather than real-world testing, the implications of this work could lead to more capable and adaptable robotic systems in the future. By improving how robots interpret and navigate their surroundings autonomously, organizations may find new ways to integrate these technologies into practical applications, enhancing operational capabilities and safety.

Academic Abstract

Humanoid parkour requires robots to reason about future footholds and whole-body motions over highly discontinuous terrain. We revisit the information asymmetry in asymmetric actor-critic learning for perceptive humanoid locomotion: the critic exploits privileged simulator signals during training, whereas the deployable actor operates solely from onboard perception. We argue that the key is not exposing privileged information to the actor, but learning a deployable approximation of it to narrow the information gap between training and deployment. We present HC2Robot, a deployable perceptive humanoid parkour framework consisting of a Predicted Foresight Module (PFM) and a Hyper-Connection (HC) Transformer. PFM predicts from deployable observations the policy-conditioned timing, location, and support quality of each foot's next touchdown event, converting privileged supervision into a deployable latent representation that provides the actor with predicted touchdown information without privileged runtime sensing. The HC Transformer further enhances temporal reasoning through learnable multi-stream routing, while a future motion prediction head provides short-horizon motion priors for regularization. Across stairs, gaps, high platforms, and slopes, HC2Robot achieves a mean simulated traversal success rate of 97.60% and consistently higher safe foothold rates than strong perceptive locomotion baselines, supporting improved traversal performance and geometric foothold support under deployable sensing.