In this video, we present KIMLAB’s recent work toward hardware-based reinforcement learning (RL) and the development of character-like bipedal robots. We explore the use of HybridLeg as a platform for expressive character robots such as Olaf or Snogie, equipped with a lantern-shaped, sensorized mechanical cover that can safely support whole-body contacts.
Humanoid robots are inherently unstable, making fall management a critical challenge, especially for on-hardware RL, where falls are frequent and unavoidable. To address this, we introduce a protective mechanical design that mitigates impact damage during falls while enabling the robot to autonomously recover to a standing posture. This capability allows the robot to self-reset and reinitialize after each trial without human intervention, which is essential for scalable real-world hardware RL. The system integrates a multimodal fall detection framework combining inertial, proprioceptive, and acoustic sensing, along with an improved stance phase detection algorithm. This work highlights a practical pathway toward robust, self-recovering humanoid platforms suitable for long-horizon, real-world reinforcement learning experiments.