Technical Summary: SPARK (Safe Protective and Assistive Robot Kit)

1. Problem Statement

Humanoid robots possess significant physical capabilities that pose substantial safety risks when interacting with complex environments and humans. While numerous safe control approaches exist—ranging from analytical methods like Control Barrier Functions (CBF) to data-driven policies—synthesizing and deploying these controllers remains inefficient for non-experts.

The core challenges identified are:

• Lack of Modularity: Synthesizing a safe controller from scratch for each new robot, task, or environment is time-consuming. Existing toolboxes often lack the modularity to seamlessly integrate with diverse frameworks or handle complex, high-dimensional systems like humanoids.
• Scalability and Generalization: A single universal safe controller capable of handling varying tasks, diverse robot dynamics, and changing environments is difficult to design. Current benchmarks often focus on safe reinforcement learning (RL) or simple mobile robots, failing to provide standardized tasks for comparing model-based safe control methods on complex humanoid platforms.
• Deployment Barriers: Efficient deployment requires easy specification of objectives and constraints, informed selection of control approaches, and tuning-free integration with real hardware, which current systems do not fully support.

2. Methodology: The SPARK Framework

To address these challenges, the authors introduce SPARK (Safe Protective and Assistive Robot Kit), a comprehensive, modular, and composable benchmark framework designed for the synthesis, testing, and deployment of safe controllers for humanoid robots.

A. Modular Architecture

SPARK decomposes the control synthesis process into distinct, interchangeable modules based on the natural dependencies of robotic control systems:

1. Configuration Module: Encapsulates robot-specific dynamics (e.g., degrees of freedom, motor interfaces) and system models.
2. Agent Module: Acts as the interface to the physical world or simulation. It handles state execution and supports both simulated agents (MuJoCo) and real robots (e.g., Unitree G1) via hardware SDKs.
3. Task Module: Aggregates system states, robot states, and user-defined goals (e.g., teleoperation commands, trajectories) into a unified data structure. It decouples information gathering from the control algorithm.
4. Algorithm Module (Back-end):
   • Policy Module: Generates nominal control commands to achieve performance objectives (e.g., reaching a goal) without considering safety.
   • Safety Module: Modifies or overrides nominal commands to enforce safety constraints. This decoupled design allows users to swap safety algorithms without altering the nominal policy.

B. Safety Options and Algorithms

SPARK implements a suite of energy-function-based safety algorithms (value-based methods) that define a scalar energy function $\phi(x)$ to quantify risk. The framework supports:

• Safe Set Algorithm (SSA): Uses a safety index to ensure the system remains within a safe set.
• Control Barrier Function (CBF): Enforces safety by ensuring $\dot{\phi} < -\alpha(\phi)$.
• Sublevel Safe Set (SSS): Combines SSA and CBF strengths, applying corrections only when constraints are active.
• Potential Field Method (PFM): Uses repulsive forces in Cartesian space.
• Sliding Mode Algorithm (SMA): Maintains the system state around a sliding layer defined by $\phi=0$.

These algorithms solve constrained optimization problems (typically Quadratic Programming) to minimize deviation from the reference control while satisfying safety constraints derived from pairwise distance functions between robot links and obstacles.

C. Benchmark Suite

SPARK provides a standardized testing suite featuring:

• Robot Configurations: Based on the Unitree G1 humanoid, including variations like fixed-base, mobile-base (floating), and sport-mode (learned policy) dynamics.
• Task Objectives: Arm manipulation (reaching 3D targets) and base navigation (2D goal reaching).
• Constraints: Static and dynamic obstacles (modeled as spheres and boxes), self-collision avoidance, and varying obstacle densities.
• Metrics: Step-wise average scores for arm/base goal tracking ($J_{arm}, J_{base}$) and self/environment safety ($M_{self}, M_{env}$).

3. Key Contributions

• A Modular Framework: SPARK offers a "plug-and-play" architecture that allows users to assemble custom safe control scenarios by swapping modules (e.g., changing robot dynamics, tasks, or safety algorithms) without re-engineering the entire system.
• Standardized Benchmarking: It provides the first standardized benchmark specifically for comparing model-based safe control approaches on high-dimensional humanoid robots, addressing a gap in existing literature which focuses heavily on safe RL or simpler platforms.
• Sim-to-Real Deployment: The framework bridges simulation and reality by supporting real-time middleware (ROS, DDS) and external sensors (Apple Vision Pro, Motion Capture), enabling the direct deployment of synthesized controllers on real hardware (Unitree G1).
• Comprehensive Algorithm Suite: It implements and compares five distinct safe control algorithms, providing a baseline for future research and a toolbox for practitioners.

4. Experimental Results

The authors evaluated SPARK through simulation benchmarks and real-robot case studies using the Unitree G1.

• Algorithm Performance:
  • Optimization-based methods (SSA, CBF, SSS) demonstrated a superior balance between safety and efficiency compared to PFM, which struggled in high-dimensional joint spaces.
  • SSS achieved the highest success rates for whole-body motion tasks, while SMA performed well for fixed-base tasks.
  • CBF excelled in environmental safety but sometimes sacrificed motion accuracy.
  • PFM consistently showed the lowest performance in complex, constrained environments.
• Trade-offs: The study highlighted the inherent trade-off between safety and efficiency. Optimization-based methods generally prioritized safety, while non-optimization methods like SMA offered better efficiency in simpler scenarios.
• Real-World Deployment:
  • Safe Autonomy: The real robot successfully avoided collisions with human hands while tracking static and dynamic targets.
  • Safe Teleoperation: Using Apple Vision Pro, the robot safely executed teleoperation commands (e.g., retrieving objects) while prioritizing collision avoidance when a human hand entered the safety zone.
  • Sim-to-Real Correlation: Statistical analysis showed high Pearson correlation ($r > 0.7$) between simulation and real-world metrics for goal tracking and safety distances, validating the framework's fidelity.
  • Computational Cost: The system maintained a runtime frequency of approximately 70–100 Hz on a standard laptop CPU, sufficient for real-time control.

5. Significance and Claims

The paper claims that SPARK significantly lowers the barrier for synthesizing and deploying safe controllers for complex robotic systems. By providing a modular, extensible, and composable framework, it enables researchers to:

• Rapidly prototype and benchmark safe control strategies across diverse robot configurations and tasks.
• Transition seamlessly from simulation to real-world deployment with minimal friction.
• Systematically evaluate the trade-offs between safety and performance, fostering the development of more robust humanoid autonomy.

The authors acknowledge limitations, noting that the current implementation does not distinguish between inevitable collisions and method failures, and that "model mismatch" between simulation and reality remains a challenge. They position SPARK as a foundational tool to accelerate research, with future work planned to integrate safe reinforcement learning, formal verification tools, and automated strategy tuning.