Project description to be revealed soon.

We present a unified whole-body safe-control framework for robotic systems that replaces nominal-plus-filter pipelines with a single quadratic program powered by Koopman neural dynamics. The method learns a Koopman embedding and globally linear dynamics from data, enabling linear optimal control and hard safety enforcement for high-dimensional, nonlinear systems within one QP. To maintain feasibility near the safe-set boundary, we introduce an adversarial fine-tuning procedure for the safety index that preserves forward invariance without degrading performance. The approach reasons over link-level, distributed safety indices and integrates cleanly with a velocity-level controller.

SPARK (Safe Protective and Assistive Robot Kit) is a modular toolbox and benchmark for humanoid autonomy and teleoperation. It integrates state-of-the-art safe control in a composable framework, making it easy to tailor protective behaviors to varied tasks, environments, and robot models. Users can set safety criteria, tune sensitivity, and compose safeguards. SPARK provides simulation benchmarks to compare methods and supports rapid deployment of synthesized controllers on real robots. It interfaces with Apple Vision Pro or Motion Capture (and other setups) and is demonstrated in simulation and on a Unitree G1 to streamline humanoid safety research.

We developed a RL environment for autonomous driving using the commercial simulator IPG CarMaker, incorporating realistic vehicle dynamics. A TD3-based control policy was trained to handle continuous steering and throttle inputs. This work demonstrates the potential of learning-based driving in high-fidelity simulation and lays the foundation for future multi-agent and complex scenario extensions.