💻 computer science

Full-Body Dynamic Safety for Robot Manipulators: 3D Poisson Safety Functions for CBF-Based Safety Filters

This paper proposes a framework for full-body dynamic collision avoidance in robotic manipulators that synthesizes a globally smooth Control Barrier Function by solving Poisson's equation on a buffered domain, thereby reducing complex multi-constraint safety requirements to a single, computationally efficient safety filter.

Original authors: Meg Wilkinson, Gilbert Bahati, Ryan M. Bena, Emily Fourney, Joel W. Burdick, Aaron D. Ames

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Meg Wilkinson, Gilbert Bahati, Ryan M. Bena, Emily Fourney, Joel W. Burdick, Aaron D. Ames

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guide a giant, clumsy octopus through a crowded room filled with moving furniture, other people, and fragile vases. The octopus has eight long, flexible arms (the robot's joints), and you need to make sure every single inch of its body avoids hitting anything, not just its "hands" (the end-effector).

This is the challenge the paper tackles: How do you keep a complex robot safe in a chaotic, moving world without slowing it down or making it freeze up?

Here is the breakdown of their solution, using simple analogies.

1. The Problem: The "Too Many Rules" Dilemma

Traditionally, to keep a robot safe, engineers try to write a rule for every single point on the robot's body.

The Analogy: Imagine trying to write a specific "Do Not Touch" rule for every single pixel on a giant, wiggly cloud. If you have 1,000 points, you have 1,000 rules.
The Issue: Computers hate having too many rules at once. If you ask a computer to solve 1,000 complex math problems simultaneously to decide where to move next, it gets overwhelmed, crashes, or takes too long. Also, if the "rules" are jagged or bumpy (mathematically speaking), the robot gets confused and jerky.

2. The Solution: The "Smooth Bubble" (Poisson Safety Functions)

The authors invented a new way to think about safety called Poisson Safety Functions (PSFs).

The Analogy: Instead of drawing a jagged line around every obstacle, imagine the obstacles are covered in a thick, invisible, smooth jelly.
- The jelly is thickest right next to the obstacle and gets thinner as you move away.
- The "safety zone" is simply the area where the jelly is thick enough.
- Because the jelly is a smooth, continuous substance (like water), the robot can easily calculate how to slide through it without getting stuck on a "jagged edge."
Why it's cool: Instead of needing a different rule for every chair, table, or person, the computer solves one big math puzzle (Poisson's equation) that creates one single, smooth safety map for the entire room. It's like having one giant, flexible bubble that expands and contracts around all the obstacles at once.

3. The Trick: The "Sampling & Buffer" Strategy

Here is the tricky part: The robot is a solid, continuous surface. You can't check every point on it (that's too much math). So, the team decided to only check a few specific "sample points" on the robot's body.

The Analogy: Imagine the robot is a fuzzy caterpillar. Instead of checking every single hair on the caterpillar, you put a few bright stickers on its back.
The Buffer: Now, imagine you shrink the "safe room" (the jelly bubble) slightly. You make the room smaller by the size of the distance between your stickers.
The Logic: If you keep the stickers inside this smaller, shrunken room, you are mathematically guaranteed that the entire fuzzy caterpillar (even the parts between the stickers) is safe in the real, larger room.
- It's like wearing a suit of armor that is slightly too big. If the armor fits inside a narrow hallway, your body definitely fits too.

4. The Result: A Real-Time "Safety Filter"

The robot runs a "Safety Filter" in real-time (like a bouncer at a club).

The robot wants to move fast (the "Nominal Controller").
The Safety Filter checks: "If I move like that, will my stickers hit the shrunken room?"
If yes, the filter gently nudges the robot's path just enough to stay safe, using the smooth "jelly" map to find the easiest way around.
If no, the robot moves exactly as planned.

5. The Real-World Test

The team tested this on a 7-armed robot (the Franka Emika) in a lab.

Static Test: They threw boxes and spheres in the way. The robot danced around them perfectly.
Dynamic Test: They had a second robot (a UR10e) with a ball attached to it, moving unpredictably into the workspace. The first robot had to dodge this moving "intruder" in real-time.
The Outcome: The robot successfully avoided collisions with the moving intruder, even when the intruder moved fast and erratically. The "stickers" (sample points) stayed safe, which meant the whole robot stayed safe.

Summary

This paper gives robots a superpower: the ability to see the whole world as one smooth, safe bubble and check only a few key points on their own body to guarantee they won't crash. It turns a mathematically impossible problem (checking infinite points) into a manageable one (checking a few points in a shrunken room), allowing robots to work safely alongside humans in busy, changing environments.

1. Problem Statement

Robotic manipulators operating in unstructured and dynamic environments face a critical challenge: ensuring full-body collision avoidance. While Control Barrier Functions (CBFs) have proven effective for safety, applying them to full-body manipulation introduces two major bottlenecks:

Differentiability and Geometry: Traditional safety representations like Signed Distance Functions (SDFs) lose differentiability in regions where the nearest obstacle is not unique (e.g., corners or complex shapes). Furthermore, approximating arbitrary obstacle geometries with primitive shapes (spheres, boxes) introduces excessive conservatism. Additionally, constructing a separate barrier function for every obstacle leads to computational complexity.
Continuous vs. Discrete Safety: Existing multi-constraint CBF approaches enforce safety only at a finite set of sampled points on the robot's surface. They lack a formal guarantee that the continuous surface between these points remains safe, potentially leading to collisions in the gaps between samples.
Scalability: Enforcing a large number of constraints in a Quadratic Program (QP) often leads to infeasibility or requires relaxation (slack variables), which weakens safety guarantees.

2. Methodology

The authors propose a unified framework combining 3D Poisson Safety Functions (PSFs) with a sampling-and-buffering strategy to achieve provable full-body safety in real-time.

A. Poisson Safety Functions (PSFs)

Instead of using SDFs or primitive approximations, the method constructs a safety function by solving Poisson's equation (an elliptic partial differential equation) on the free space defined by an occupancy map.

Formulation: Given a free space domain $\Omega$ with boundary $\partial\Omega$ (obstacle surfaces), the PSF $h_0$ is the unique solution to:
$\nabla^2 h_0(y) = f(y) \quad \text{for } y \in \Omega, \quad h_0(y) = 0 \quad \text{for } y \in \partial\Omega$
where $f(y) < 0$ is a forcing function.
Advantages:
- Global Smoothness: By elliptic regularity, the solution $h_0$ is infinitely smooth ( $C^\infty$ ) everywhere in the interior, ensuring well-defined gradients for the CBF-QP.
- Arbitrary Geometry: It handles complex, non-convex obstacle shapes without over-approximation.
- Single Function: A single PDE solve generates one global safety function for the entire environment, avoiding the need for multiple barrier constructions.

B. Sampling and Buffering for Full-Body Safety

To bridge the gap between discrete sampling and continuous surface safety, the authors introduce a geometric proof based on the Pontryagin difference.

Sampling: The robot surface is approximated by a set of $N$ sample points $Y = \{y_1, \dots, y_N\}$ such that every point on the continuous surface is within a distance $\epsilon$ of a sample point (coverage condition).
Buffering: The safe region $C$ (free space) is shrunk by $\epsilon$ using the Pontryagin difference: $C_\epsilon = C \ominus B_\epsilon$ .
Theorem: If all sample points $y_i$ $y_{i}$ are kept within the buffered region $C_\epsilon$ $C_{ϵ}$ , then the entire continuous robot surface $S(q)$ $S (q)$ is guaranteed to be within the true safe region $C$ $C$ .
- Logic: If a sample point is at least $\epsilon$ away from the obstacle boundary, the $\epsilon$ -ball around it (covering the unsampled surface area) cannot intersect the obstacle.

C. Real-Time Safety Filter

The system synthesizes a time-varying PSF $h_\epsilon(y, t)$ for dynamic environments. This function is evaluated at the $N$ sample points to generate a Multi-Constraint CBF-QP:

Objective: Minimize deviation from a nominal velocity command ( $v_{nom}$ ).
Constraints: Enforce $\dot{h}_\epsilon(y_i, t, v) \geq -\alpha(h_\epsilon(y_i, t))$ for all $N$ sample points simultaneously.
Robustness: An adaptive robustness term is added to the CBF constraint to handle sensor latency and encoder errors (Input-to-State Safe CBF).

3. Key Contributions

PSF-Based CBF Filter: Introduction of a single, globally smooth safety function derived from Poisson's equation, eliminating the need for ad-hoc geometric primitives and ensuring differentiability for QP solvers.
Formal Full-Body Guarantee: A mathematical proof demonstrating that enforcing safety on a discretely sampled set of points, combined with a Pontryagin difference buffer, guarantees collision avoidance for the entire continuous robot surface.
Hardware Validation: Successful implementation on a 7-DOF Franka Emika FR3 manipulator in both static and dynamic environments, demonstrating real-time performance (50–100 Hz).

4. Experimental Results

The framework was validated on a 7-DOF robot with the following results:

Static Environments: The robot successfully navigated cluttered environments with adversarial commands (intentional attempts to collide). The safety filter prevented collisions while maintaining high motion fidelity (not overly conservative).
Dynamic Environments:
- A second robot (UR10e) carrying a spherical obstacle was teleoperated into the workspace of the FR3.
- The occupancy grid was updated in real-time (100 Hz), and the PSF was recomputed online.
- The FR3 successfully avoided collisions with the moving obstacle, even when the obstacle's motion was unpredictable and un-planned.
Performance Metrics:
- PDE Solve Time: ~0.002s (using Successive Over-Relaxation).
- QP Solve Time: ~0.003s.
- Sampling Resolution: A resolution of $\epsilon = 0.1$ resulted in ~30 sample points, balancing safety and computational load.
- Robustness: The system maintained strictly positive safety function values ( $h > 0$ ) across all sample points throughout the experiments.

5. Significance

This work addresses a fundamental limitation in robotic safety: the trade-off between computational tractability and provable full-body safety.

Theoretical Impact: It provides a rigorous proof that discrete sampling + buffering is sufficient for continuous safety, removing the "gap" in existing multi-constraint CBF methods.
Practical Impact: By using a single smooth PSF, the method avoids the infeasibility issues common in multi-obstacle CBF formulations and eliminates the conservatism of primitive-based approximations.
Dynamic Capability: The ability to synthesize safety functions online from occupancy maps enables safe operation in unknown, dynamic environments where obstacles move unpredictably, a critical requirement for human-robot collaboration.

In summary, the paper presents a robust, mathematically grounded, and computationally efficient framework that enables robot manipulators to operate safely in complex, dynamic 3D environments without sacrificing performance or safety guarantees.