💻 computer science

Surface Constraint Policy for Learning Surface-Constrained and Dynamically Feasible Robot Skills

This paper proposes a Surface Constraint Policy (SCP) that integrates a two-dimensional weighted Gaussian kernel for encoding free-form surface geometry with a diffusion-based model and dynamic movement primitives to generate robot actions that ensure reliable surface alignment and dynamic feasibility during complex manipulation tasks.

Original authors: Shuai Ke, Jiexin Zhang, Huan Zhao, Zhiao Wei, Yikun Guo, Jie Pan, Han Ding

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuai Ke, Jiexin Zhang, Huan Zhao, Zhiao Wei, Yikun Guo, Jie Pan, Han Ding

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 teaching a robot to clean a surface, like wiping a whiteboard or polishing an airplane window. If the surface is perfectly flat, it's easy. But what if the surface is curved, bumpy, or shaped like a complex airplane window? That's where things get tricky.

This paper introduces a new "brain" for robots called the Surface Constraint Policy (SCP). Think of it as a smart guide that helps a robot learn to clean complex shapes without losing its grip or scratching the surface.

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

The Problem: The "Clumsy Robot"

Current robot learning methods are like students who memorize a dance routine but forget the rules of the dance floor.

The Issue: When robots try to learn from human demonstrations (like watching a video of someone wiping a window), they often guess the next move. Because they don't have a strict map of the surface's shape, they might guess a move that lifts the tool off the glass or hits it at a weird angle.
The Result: The robot loses contact, leaves streaks, or moves so jerkily that the motion is unstable. It's like trying to walk a tightrope while blindfolded; you might get there, but you'll probably fall off.

The Solution: The "Surface Constraint Policy" (SCP)

The authors built a system that gives the robot two superpowers: Geometry Awareness and Smooth Movement.

1. Mapping the Invisible Road (Geometry Encoding)

Before the robot even starts moving, the system takes the human's demonstration and turns the 3D shape of the surface into a mathematical "map."

The Analogy: Imagine you are drawing a path on a piece of paper. Now, imagine that paper is actually a bumpy, curved hill. The system uses a special "weighted Gaussian kernel" (think of it as a smart rubber stamp) to press the human's path onto the hill. It creates a strict rule: "You must stay on this line, no matter how the hill curves."
Why it matters: This ensures the robot knows exactly where the surface is, so it doesn't accidentally float above it or dig into it.

2. The "Dreaming" Brain (Diffusion Policy)

The robot uses a "Diffusion Policy" to figure out what to do next.

The Analogy: Think of this like a sculptor starting with a block of noisy, static-filled clay. The robot starts with a rough, messy idea of the next move. Then, step-by-step, it "denoises" the clay, refining the messy idea into a clear, smooth shape. It looks at what the human did and what the camera sees right now to guess the best next move.
The Catch: Even though this "dreaming" process is great at guessing, the raw guesses can still be a bit jittery or unrealistic for a real robot arm.

3. The "Smooth Translator" (Similarity Mapping)

This is the paper's secret sauce. The robot takes the "jittery" guess from the dreaming brain and translates it into a smooth, physically possible motion.

The Analogy: Imagine the robot's guess is a rough sketch of a road. The "Similarity Mapping" is like a professional cartographer who redraws that sketch onto a real, paved highway. It ensures the road follows the curves of the terrain (the surface) and that the car (the robot arm) doesn't have to make impossible sharp turns or sudden stops.
The Result: The robot moves smoothly, keeping constant contact with the surface, just like a human would.

The Test Drive: Three Cleaning Challenges

The researchers tested this system on three tasks, getting progressively harder:

Whiteboard Wiping: A flat surface. The robot did well, but so did other methods.
Free-Form Surface Wiping: A bumpy, curved surface. Here, other robots started to fail, losing contact or moving too erratically. The SCP robot stayed on track and cleaned perfectly.
Airplane Windscreen Cleaning: The hardest test. The window is a complex, curved shape, and the tool (a squeegee) is rigid. If the angle is wrong even a little, the cleaning fails.
- The Outcome: Other methods failed often or moved dangerously. The SCP robot succeeded 98% of the time, matching human performance almost perfectly. It kept the squeegee pressed firmly against the glass at the right angle, leaving no streaks.

The Bottom Line

The paper claims that by combining a "smart map" of the surface with a "smooth translator" for the robot's movements, they created a robot that can learn complex cleaning tasks from humans and execute them with high precision. It doesn't just guess; it respects the shape of the world it's working in.

In short: They taught the robot not just what to do, but how to stay glued to the surface while doing it, resulting in a cleaner, safer, and more reliable robot.

Technical Summary: Surface Constraint Policy for Learning Surface-Constrained and Dynamically Feasible Robot Skills

Problem Statement

Diffusion-based imitation learning has advanced robot dexterous manipulation but faces significant limitations when applied to tasks involving complex free-form surface constraints (e.g., wiping, polishing, ultrasound scanning). Current generative models lack explicit surface geometry constraint modeling and often fail to ensure dynamic feasibility. This results in:

Stochastic action generation that fails to achieve reliable surface alignment.
Unstable contact due to insufficient consideration of kinematic and dynamic constraints, leading to large motion accelerations, irregular trajectory curvature, and potential loss of contact or surface damage.
Existing methods either rely on implicit learning of geometric relationships (which is imprecise) or introduce constraints during training without guaranteeing smooth, dynamically admissible execution on real hardware.

Methodology: The Surface Constraint Policy (SCP)

The authors propose a novel framework, the Surface Constraint Policy (SCP), which integrates human demonstrations and real-time visual observations to generate actions that satisfy both surface geometry constraints and dynamic feasibility. The pipeline consists of three core components:

1. Surface Geometry Encoding

To provide geometric priors, the system encodes the free-form surface $S(u, v)$ using a two-dimensional weighted Gaussian kernel function derived from demonstration trajectories.

Process: Demonstration end-effector poses are parameterized into local $uv$ coordinates.
Model: A continuous surface state function is constructed using a weighted sum of Gaussian kernels, where weights are learned via Locally Weighted Regression (LWR). This creates a structured geometric representation of the surface.

2. Diffusion-Based Action Intent Inference

A conditional denoising diffusion probabilistic model (DDPM) is used to infer task-level action intentions ( $A_t$ ) from multimodal inputs.

Inputs: RGB image sequences ( $O_t$ ), end-effector pose ( $P_t$ ), and joint configurations ( $J_t$ ).
Architecture: Visual features are encoded via an untrained ResNet-18, while proprioceptive states are encoded via an MLP. These are concatenated and processed by a transformer-based diffusion model.
Output: The model generates a sequence of future action intentions (positions and orientations) conditioned on the current observation and the diffusion step.

3. Similarity-Based Motion Mapping and Execution

To bridge the gap between stochastic policy outputs and physically executable motions, the inferred action intentions are transformed into Surface-Constrained Dynamic Movement Primitives (DMPs) via a similarity-based mapping method.

Position Mapping: The discrete trajectory points from the diffusion policy are projected onto the encoded surface $S(u, v)$ $S (u, v)$ . This is formulated as a joint optimization problem minimizing:
1. Euclidean distance between projected and original points (geometric consistency).
2. Kullback–Leibler (KL) divergence between local direction entropies before and after projection (preserving trajectory structure).
Orientation Estimation: Orientations are estimated using Gaussian intrinsic clustering on the unit quaternion manifold ( $S^3$ ), weighted by spatial distance to handle non-uniform demonstration distributions.
Dynamic Execution: The mapped trajectories are executed using:
- Arc-Length DMPs (AL-DMPs): For position trajectories, decoupling spatial shape from temporal execution speed.
- Geodesic DMPs (Geo-DMPs): For orientation trajectories, evolving along geodesics on the quaternion manifold.
- Synchronization: Both DMP types use geometric length (arc length or geodesic length) as the phase variable, ensuring consistent position-orientation coordination under variable speeds.

Key Contributions

Surface Constraint Policy (SCP) Framework: A novel policy learning framework that generates fine-grained, geometry-consistent action intents under surface constraints, addressing the lack of explicit modeling in existing diffusion policies.
Similarity Mapping Method: A mechanism to transform high-level policy intentions into dynamically feasible robot actions. This method ensures surface compliance while preserving trajectory structure, significantly improving motion smoothness and stability.
Comprehensive Validation: The method is evaluated on three tasks with increasing complexity (whiteboard wiping, free-form surface wiping, and aircraft windscreen cleaning), demonstrating superior performance over state-of-the-art baselines.

Experimental Results

The SCP was compared against three baselines:

DP: Standard diffusion policy (no explicit constraints).
MDP: Diffusion combined with movement primitives (improves smoothness but lacks explicit surface constraints).
ACP: Adaptive Compliance Policy (uses impedance control for contact force regulation).

Key Findings:

Task Success Rates: SCP achieved success rates comparable to human demonstrations (approx. 98–99%) across all tasks.
- Whiteboard Wiping: SCP (0.99) outperformed DP (0.59) and MDP (0.92).
- Free-Form Wiping: SCP (0.98) significantly outperformed MDP (0.52).
- Windscreen Cleaning: SCP (0.98) significantly outperformed ACP (0.63). DP and MDP failed consistently in this high-precision task.
Surface-Alignment Error (SAE): SCP demonstrated the lowest SAE in position and orientation, particularly in the windscreen cleaning task where rigid tool precision is critical. ACP showed higher orientation errors due to posture deviations.
Dynamic Feasibility: SCP produced smoother trajectories with lower motion acceleration and more regular curvature compared to DP and MDP, preventing contact loss and ensuring stable execution.

Significance and Claims

The paper claims that the SCP effectively addresses the dual challenges of geometric intent generation and dynamic feasibility in surface-constrained manipulation.

Robustness: By explicitly encoding surface geometry and using a similarity-based mapping to DMPs, SCP avoids the common failure modes of inaccurate geometric inference and unstable execution found in purely generative approaches.
Generalizability: The method demonstrates strong generalization across tasks ranging from simple planar wiping to complex, high-curvature aircraft windscreen cleaning with rigid tools.
Limitations and Future Work: The authors note that the current method relies heavily on high-quality demonstration data. Future work aims to incorporate prior knowledge from large language models and video-language models to enable rapid adaptation to novel geometries and tools with fewer demonstrations.

The authors conclude that SCP provides a promising approach for skill learning in complex, contact-rich industrial scenarios, achieving a unified formulation of structured surface-geometry intent and dynamically feasible action execution.