SafeGuard CBF-QP: VLA Safety Filter Demo

Real-time safety filtering for Vision-Language-Action (VLA) models using Control Barrier Functions.

When a VLA model hallucinates, the robot arm doesn't know. This demo shows how a CBF-QP safety layer can intercept and minimally correct unsafe trajectories before they reach the hardware — at 1kHz, in real-time.

---

What This Demo Does

1. Simulates VLA trajectories — including normal trajectories and "hallucinated" ones (OOD drift, cross-modal misalignment)
2. Applies CBF-QP safety filtering — solves a Quadratic Program at each timestep to find the minimum correction that keeps the system safe
3. Visualizes the result — shows original (unsafe) vs. corrected (safe) trajectories with ISO 15066 human body risk zones

Core Math

Control Barrier Function (CBF)

For a safety constraint $h(\mathbf{x}) \geq 0$ (e.g., "stay away from the human head"), the CBF condition requires:

$$\dot{h}(\mathbf{x}, \mathbf{u}) + \alpha \cdot h(\mathbf{x}) \geq 0$$

where $\alpha > 0$ controls how aggressively the system enforces safety.

QP Safety Filter

Given the VLA's desired control $\mathbf{u}_{VLA}$, find the closest safe control:

$$\min_{\mathbf{u}} \quad \frac{1}{2} |\mathbf{u} - \mathbf{u}_{VLA}|^2$$

$$\text{s.t.} \quad \dot{h}_i(\mathbf{x}, \mathbf{u}) + \alpha_i \cdot h_i(\mathbf{x}) \geq 0, \quad \forall i$$

This is a convex QP — solvable in <1ms on embedded hardware.

Quick Start

# Clone
git clone https://github.com/YOUR_USERNAME/safeguard-cbf-demo.git
cd safeguard-cbf-demo

# Install dependencies
pip install -r requirements.txt

# Run the demo
python safeguard_cbf_demo.py

# Or run the interactive notebook
jupyter notebook safeguard_demo.ipynb

Requirements

numpy>=1.21
scipy>=1.7
matplotlib>=3.5
cvxpy>=1.3        # For QP solving

What You'll See

The demo generates 4 figures:

Figure	Description
Trajectory Comparison	VLA hallucinated trajectory vs. CBF-corrected trajectory
Safety Margin	$h(\mathbf{x})$ over time — stays positive after filtering
Control Effort	How much correction was needed at each timestep
ISO 15066 Risk Map	Human body part zones with different force limits

Architecture

VLA Model Output (u_vla)
        │
        ▼
┌─────────────────┐
│  CBF-QP Filter   │  ← Solves QP at each timestep
│  (this demo)     │  ← Multiple safety constraints
│                  │  ← ISO 15066 body part mapping
└────────┬────────┘
         │
         ▼
  Safe Control (u_safe)
         │
         ▼
  Robot Hardware

Related Patents

This demo is a simplified illustration of the engineering principles behind:

• G2: CBF-based VLA hallucination interception with ISO 15066 mapping
• L3: Cross-modal consistency verification for hallucination detection
• M1: Conformal prediction-based dynamic threshold calibration

Part of a 71-patent portfolio for embodied AI safety. Contact us for the full SafeGuard SDK.

License

MIT — use freely for research and development.

Contributing

Issues and PRs welcome! Areas where we'd love help:

• Adapting to ROS2/MoveIt2
• Franka Emika / UR5e real-hardware integration
• Additional safety constraint types