Safe-SAGE: Social-Semantic Adaptive Guidance for Safe Engagement through Laplace-Modulated Poisson Safety Functions

Safe-SAGE is a unified framework that bridges high-level semantic understanding with low-level safety-critical control by employing a Laplace-modulated Poisson safety function within a multi-layer filter, enabling legged robots to navigate dynamic, semantically rich environments with context-dependent safety margins while maintaining rigorous guarantees.

The Gist

Imagine you are walking down a busy hallway. You see a person walking toward you, a stack of cardboard boxes, and a sturdy brick wall.

A traditional robot, using old-school safety software, sees all three of these things exactly the same way: "Obstacle. Stop or turn." It treats the person, the boxes, and the wall as identical geometric shapes. If you were to program a robot to be safe, it might give the person the same wide berth as the wall, or worse, it might be too aggressive and bump into the person because it doesn't understand that a human is fragile and expects you to step aside politely.

This paper introduces a new system called Safe-SAGE. Think of it as giving the robot a "social brain" and a "safety instinct" that work together.

Here is how it works, broken down into simple concepts:

1. The "Social Radar" (Perception)

First, the robot needs to know what it is looking at, not just where it is.

• Old Way: The robot sees a blob of pixels.
• Safe-SAGE Way: The robot uses its cameras and lasers like a super-powered pair of eyes. It doesn't just see "stuff"; it identifies "That's a human," "That's a chair," and "That's a wall." It even keeps track of people even if they walk behind a corner where the camera can't see them, kind of like how you remember your friend is still in the room even if they step behind a door.

2. The "Invisible Traffic Cop" (The Guidance Field)

Once the robot knows what the objects are, it creates an invisible map of how it should behave. The authors call this a Laplace Guidance Field.

• The Metaphor: Imagine the robot is a boat in a river.
  • The Wall: The river current pushes the boat hard away from the wall. It's a "no-go" zone.
  • The Boxes: The current pushes the boat away, but gently. You can brush past them if you have to.
  • The Human: This is where the magic happens. The current doesn't just push the boat away; it creates a rotational flow. It gently nudges the boat to pass the human on their left (a common social rule, like driving on the right side of the road). It creates a "personal space bubble" that is much larger for the human than for the boxes.

This "current" is generated by a math equation (Poisson Safety Function) that acts like a pressure cooker. The pressure is highest near humans and lower near inanimate objects, forcing the robot to respect social norms automatically.

3. The "Double-Check Safety Net" (The Filter)

The robot has two layers of safety checks to make sure it never crashes, even if the "traffic cop" gets confused.

• Layer 1: The Crystal Ball (MPC): This is the robot's planner. It looks a few seconds into the future and asks, "If I go this way, will I hit someone?" It plans a smooth path that respects the social rules.
• Layer 2: The Reflex (Real-time Filter): This is the robot's knee-jerk reaction. If a human suddenly jumps out from behind a corner, the Crystal Ball might be too slow to react. The Reflex layer acts like a brake pedal. It instantly calculates, "Wait, that's a human! Stop or swerve immediately!" It overrides the plan to ensure safety in the split second.

Why is this a big deal?

Before this, robots had to choose between being too scared (stopping at everything, making them useless in crowds) or too reckless (ignoring social cues and scaring people).

Safe-SAGE allows the robot to be smartly cautious.

• It knows a wall is hard, so it gives it a little space.
• It knows a human is soft and has feelings, so it gives them a lot of space and passes them politely.
• It knows a chair is just an object, so it can squeeze past it if the hallway is narrow.

The Bottom Line

The researchers tested this on a four-legged robot (like a dog) and even a humanoid robot. They found that the robots could walk through crowded cafeterias and hallways, avoiding collisions while acting like polite humans. They didn't just avoid crashing; they followed the unwritten rules of society, like walking on the left side of a person.

In short, Safe-SAGE teaches robots to stop thinking of the world as a collection of geometric shapes and start seeing it as a social environment where different things require different kinds of respect.

Technical Summary

Here is a detailed technical summary of the paper "Safe-SAGE: Social-Semantic Adaptive Guidance for Safe Engagement through Laplace-Modulated Poisson Safety Functions."

1. Problem Statement

Traditional safety-critical control methods (e.g., Control Barrier Functions - CBFs, Model Predictive Control - MPC) often suffer from "semantic blindness." They treat all obstacles geometrically, meaning a robot applies the same avoidance behavior to a human, a chair, or a wall if they occupy the same geometric volume. This leads to two suboptimal outcomes:

• Universal Conservatism: Treating all objects as high-risk degrades performance in cluttered spaces.
• Universal Aggression: Treating all objects as low-risk risks safety failures.

Furthermore, while Large Language Models (LLMs) and Vision-Language Models (VLMs) can understand semantics, they operate at low frequencies with high latency, making them unsuitable for the near-real-time requirements of dynamic safety filters. There is a critical "neuro-symbolic gap" between high-level semantic reasoning and low-level safety enforcement.

2. Methodology: Safe-SAGE Framework

The authors propose Safe-SAGE, a unified framework that bridges this gap by embedding social and semantic understanding directly into the safety filter layer using Poisson Safety Functions (PSFs) modulated by Laplace Guidance Fields (LGFs).

A. System Architecture

The system operates as a layered architecture between perception and control:

1. Perception & Tracking: Fuses multi-sensor point clouds (LiDAR) and RGB images. It uses instance segmentation (YOLOv11) for semantic labeling and an object-level tracker to maintain persistent human tracking even when humans move out of the camera's field of view.
2. Semantic Scene Understanding: Generates a semantic occupancy grid where cells are labeled (e.g., Human, Wall, Chair).
3. Field Synthesis:
   • Laplace Guidance Field (LGF): Constructs a vector field $\mathbf{v}$ that encodes social norms. It solves a vector Laplace equation with class-aware boundary conditions:
     • Normal Flux: Repulsion magnitude ($b(q)$) varies by object class (e.g., stronger repulsion for humans than walls).
     • Tangential Bias: Introduces a rotational component to enforce social passing norms (e.g., passing on the left or right).
   • Poisson Safety Function (PSF): Uses the divergence of the LGF as a forcing term to solve Poisson's equation ($\Delta h = \nabla \cdot \mathbf{v}$). This generates a scalar safety function $h$ where the "safety margin" steepness varies based on semantic importance.
4. Dual-Layer Safety Filter:
   • Predictive Layer (MPC): Plans optimal trajectories over a finite horizon subject to linearized CBF constraints derived from the semantic LGF.
   • Reactive Layer (Analytical CBF): A real-time filter that projects nominal control inputs onto the safe set using a closed-form solution. It includes a temporal derivative term ($\partial h / \partial t$) to account for dynamic obstacles and motion compensation.

B. Key Technical Mechanisms

• Semantic Flux Modulation: The boundary conditions of the Laplace equation are tuned per object class. For example, $b_{human} = -1.7$ (strong repulsion) vs. $b_{wall} = -0.5$ (weak repulsion).
• Social Bias Integration: A tangential bias is added to the LGF on an internal Dirichlet interface, creating a rotational flow that guides the robot to pass humans on a specific side (e.g., left), satisfying social norms without explicit path planning.
• Forward Invariance: The paper provides a formal proof (Theorem 1) that the modified guidance field maintains the forward invariance of the safe set, ensuring rigorous safety guarantees despite the non-conservative nature of the rotational field.

3. Key Contributions

1. Social-Biased Guidance Field: A novel formulation embedding social biases (directional avoidance norms) directly into the Laplace guidance field, enabling context-dependent safety margins.
2. Formal Safety Analysis: A rigorous proof demonstrating that the modified, rotationally-biased guidance field preserves the forward invariance of the safe set, ensuring safety guarantees are not compromised by social behaviors.
3. Integrated Pipeline: A complete pipeline for real-time integration of class-aware inflation and semantic flux modulation, generating obstacle-conforming safety fields from raw sensor data.

4. Experimental Results

The framework was validated on a Unitree Go2 quadruped robot and a Unitree G1 humanoid robot in various environments (hallways, open areas, crowded cafeterias).

• Simulation & Hardware Benchmarks:
  • Social Compliance: The robot successfully navigated gaps between humans and static obstacles, maintaining a wider margin from humans than from walls.
  • Passing Norms: The robot consistently passed humans on the designated side (e.g., left), whereas baseline methods without semantic bias passed randomly or collided.
• Quantitative Metrics (Table I):
  • Human-Robot Margin: The proposed method achieved a positive margin of 0.318m (keeping a safe distance), while the baseline resulted in a negative margin (-0.008m), indicating collision or near-collision.
  • Max Lateral Offset: The proposed method achieved 0.75m lateral offset (successfully passing), compared to -0.1m for the baseline (failing to pass).
• Generalization: The system demonstrated robustness in dynamic environments with sensory noise and successfully transferred to different hardware platforms (quadruped vs. humanoid) with minimal modification.

5. Significance

Safe-SAGE represents a significant advancement in robotic safety by moving beyond purely geometric constraints to semantic-aware safety.

• Bridging the Latency Gap: It avoids the latency issues of LLMs/VLMs by distilling semantic understanding into mathematical safety functions that operate at control frequencies (high Hz).
• Context-Aware Conservatism: It allows robots to be "conservative" with humans and "aggressive" (efficient) with static furniture, optimizing performance without sacrificing safety.
• Social Navigation: It provides a mathematically rigorous method for enforcing social norms (like passing sides) within safety-critical control loops, a capability previously difficult to achieve with hard safety guarantees.

In summary, Safe-SAGE enables legged robots to navigate complex, human-centric environments with the same level of safety rigor as traditional methods, but with the added intelligence to distinguish between a wall and a person, adapting their behavior accordingly.