💻 computer science

Stage-Aware and Roughness-Constrained Diffusion Policy for Multi-Stage Robotic Polishing

This paper proposes a Stage-Aware and Roughness-Constrained Diffusion Policy (SRDP) that leverages multimodal observations to infer process stages and enforce physical constraints on feed speed and contact force, thereby enhancing stage-transition stability and surface quality in multi-stage robotic polishing tasks.

Original authors: Shuai Ke, Jiexin Zhang, Huan Zhao, Zhiao Wei, Yikun Guo, Tiange Wu, Guoqiang Guo, Haoyuan Zhou, Jie Pan, Han Ding

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuai Ke, Jiexin Zhang, Huan Zhao, Zhiao Wei, Yikun Guo, Tiange Wu, Guoqiang Guo, Haoyuan Zhou, 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 be a master craftsman, specifically one that polishes the inside of a spacecraft. This isn't just about moving a hand from point A to point B; it's a complex, multi-step dance where the robot must switch between different "modes" (like polishing a flat surface, then cleaning up dust, then polishing a curved edge) while keeping its touch perfectly gentle and consistent.

The paper introduces a new "brain" for the robot called SRDP (Stage-Aware and Roughness-Constrained Diffusion Policy). Here is how it works, broken down into simple concepts:

1. The Problem: The Robot Gets Confused and Drifts

The authors identified two main headaches when trying to teach robots these complex tasks:

The "Where Am I?" Problem (Stage Uncertainty): Imagine you are walking through a house where every room looks exactly the same. You might forget if you are in the kitchen or the bedroom. Similarly, a robot looking at a spacecraft part might see a similar-looking surface but not know if it should be polishing, cleaning, or just holding the object. Existing robots often get stuck or switch to the wrong action because they can't tell which stage of the process they are in.
The "Drifting" Problem (Parameter Drift): Imagine a chef trying to bake a cake. If they guess the amount of sugar and flour without a recipe, they might accidentally mix in too much sugar or too little flour. In polishing, the robot controls how fast it moves (feed speed) and how hard it presses (force). If these two numbers drift apart randomly, the robot might scratch the surface (pressing too hard) or leave it rough (pressing too lightly).

2. The Solution: A "Smart Guide" and a "Safety Net"

The SRDP framework solves these problems with two clever tricks:

Trick A: The "Memory Lane" (Stage-Awareness)

Instead of just looking at the current picture, the robot looks at its recent history (what it saw and did in the last few seconds).

The Analogy: Think of a detective solving a mystery. A detective doesn't just look at a single clue; they look at the sequence of events to figure out where they are in the story.
How it works: The robot uses a "Stage Inference Network" to guess, "Okay, based on what I just did, I am probably in the 'Polishing the Edge' stage, not the 'Cleaning' stage." It then uses this guess to tell the rest of its brain exactly what kind of moves to make next. This keeps the robot on the right "track" of the workflow.

Trick B: The "Quality Filter" (Roughness Constraints)

The robot doesn't just guess how hard to press; it has a built-in rulebook.

The Analogy: Imagine a musician playing a song. They don't just hit random notes; they follow a sheet music that ensures the notes fit together to make a harmonious chord.
How it works: The robot knows that for a specific type of polish (the "stage"), the speed of the tool and the pressure must follow a specific mathematical relationship to get a smooth finish. As the robot "dreams up" (generates) a plan for its next moves, this rulebook acts as a filter. If the plan suggests pressing too hard for the speed, the system nudges the plan back into the "safe zone" before the robot actually moves. This prevents the "drifting" problem.

3. The "Diffusion" Part: How the Robot Learns

The paper uses a technique called Diffusion Policy.

The Analogy: Imagine a sculpture made of clay that is covered in static noise (like TV snow). The robot's job is to slowly wipe away the noise, step by step, revealing the perfect sculpture underneath.
How it works: The robot starts with a chaotic, random set of movements. It then slowly "denoises" this chaos, refining it over and over until it becomes a smooth, perfect sequence of actions that matches what a human expert would do. The "Stage-Aware" and "Roughness-Constrained" parts are just special instructions given to the robot while it wipes away the noise, ensuring the final sculpture looks exactly right.

4. The Results: Real-World Testing

The team tested this on a real dual-arm robot polishing a spacecraft cabin. They compared their new method (SRDP) against other advanced robot learning methods.

The Outcome: The SRDP robot was much better at switching between tasks (like going from polishing to vacuuming) without getting confused. It also produced a much smoother, more consistent surface finish because it didn't let its pressure and speed drift apart.
The Proof: When they measured the surface roughness, the SRDP robot's work was more uniform and higher quality than the other methods.

Summary

In short, this paper presents a robot brain that is better at knowing where it is in a long process and better at keeping its physical actions (speed and pressure) in perfect harmony. By combining a "memory" of the process stages with a "rulebook" for surface quality, the robot can polish complex spacecraft parts with a level of consistency and skill that previous methods couldn't achieve.

Technical Summary: Stage-Aware and Roughness-Constrained Diffusion Policy for Multi-Stage Robotic Polishing

1. Problem Statement

Robotic polishing is a critical finishing process in high-value manufacturing sectors like aerospace, where surface quality directly dictates component reliability. However, automating these tasks presents two primary challenges that existing imitation learning methods struggle to address simultaneously:

Long-Horizon Process-Stage Uncertainty: Polishing workflows consist of multiple distinct stages (e.g., stabilization, edge polishing, surface finishing, cleaning) that often exhibit visually similar local states but require vastly different control objectives and action patterns. Existing methods often rely on explicit task decomposition or high-level planning prompts, which may not suffice for inferring latent process stages from multimodal histories in real-time.
Unconstrained Process-Parameter Drift: In diffusion-based policies, the stochastic nature of the denoising process can lead to unconstrained variations in coupled process parameters (feed speed and normal contact force). Since these parameters jointly determine material removal and surface roughness, such drift can result in physically infeasible combinations, unstable contact, or inconsistent surface quality (e.g., under-polishing or edge cutting), especially when spindle speeds are preset per stage.

Current literature lacks a unified framework that jointly handles latent stage inference from observation histories and enforces roughness-driven physical constraints within high-dimensional generative policies.

2. Methodology: The SRDP Framework

The authors propose the Stage-Aware and Roughness-Constrained Diffusion Policy (SRDP), a unified framework designed to learn multi-stage polishing skills from demonstrations while adhering to process constraints. The framework operates through two interconnected mechanisms:

A. Stage-Conditioned Action Diffusion
To address stage uncertainty, SRDP introduces a stage inference network that processes multimodal observation histories (visual images, robot states, contact forces, and tool modes) to infer a posterior probability distribution over the current process stage.

Stage Inference: A network $g_\phi$ maps the observation history $O_t$ to a stage posterior $z_t$ .
Conditioned Denoising: This posterior is encoded into a stage condition vector $e_t$ , which is injected into the shared reverse denoising network. This allows the policy to generate action sequences consistent with the inferred current stage without requiring external stage labels during execution.
Transition Consistency: To ensure smooth transitions between stages, a stage-transition consistency prior is introduced. This uses a transition matrix $T$ (derived from process flow knowledge or data) to regularize the stage inference, penalizing abrupt or illogical stage changes via a soft-label cross-entropy loss.

B. Roughness-Constrained Diffusion Sampling
To address parameter drift, SRDP embeds physical and quality constraints directly into the diffusion sampling process.

Roughness-Oriented Modeling: A power-law empirical model predicts surface roughness ( $R_a$ ) based on normal force ( $F_n$ ), feed speed ( $v$ ), and stage-wise preset spindle speed ( $\omega$ ). The model is calibrated offline using experimental data.
Constraint Sets: Two sets are defined for the parameter space:
1. Roughness-Consistency Set ( $\Omega_R$ ): Ensures the coupled relationship between $F_n$ and $v$ satisfies the target roughness within a tolerance band.
2. Physical-Feasibility Set ( $\Omega_P$ ): Enforces bounds on force/speed limits and temporal continuity (rate of change) to ensure robot executability.
Guided Sampling: During the reverse denoising steps, the policy minimizes a constraint energy function composed of roughness-consistency energy and physical-feasibility energy. A "roughness-dominant" soft guidance strategy is applied in early/mid-noise stages to steer samples toward the feasible region, followed by a minimum-modification projection in the final steps to guarantee the output lies strictly within the joint feasible set $\Omega_t = \Omega_R \cap \Omega_P$ .

3. Key Contributions

The paper claims the following contributions:

Unified Framework: The proposal of SRDP, which integrates stage inference from multimodal histories with shared reverse denoising to enable stage-consistent action generation without external stage labels.
Process-Constrained Sampling: The development of a roughness-oriented constraint-guided diffusion sampling method that generates feed speed and normal contact force satisfying both surface quality requirements and physical feasibility under preset spindle speeds.
Empirical Validation: Comprehensive evaluation through comparative experiments, ablation studies, and real-robot validations in two complex scenarios: spacecraft cabin coating-surface polishing and inner-cavity structural surface finishing.

4. Experimental Results

Experiments were conducted on a dual-arm robotic platform (two KUKA iiwa14 robots) performing two distinct tasks:

Coating-Surface Polishing: Involving stabilization, square/circular hole polishing, and vacuum cleaning.
Inner-Cavity Finishing: Involving cooperative transport, stabilization, chamfer refinement, and assembly-surface polishing.

Performance Metrics:

Task Success: SRDP achieved the highest average subtask success rates (0.94 for coating, 0.91 for inner-cavity) and transition success rates (0.88 and 0.83, respectively), significantly outperforming baselines like standard Diffusion Policy (DP), Hierarchical Imitation Learning (HIL), and Adaptive Compliance Policy (ACP).
Failure Analysis: SRDP substantially reduced the two dominant failure modes:
- Stage-Switching Failures (STF): Reduced to ~21-24% compared to >50% in baselines without explicit stage modeling.
- Process-Parameter Failures (PPF): Reduced to ~9.5-10.5% compared to higher rates in methods lacking roughness constraints.
Surface Quality: SRDP demonstrated lower and more stable surface roughness values and more consistent chamfer widths compared to ACP, validating the effectiveness of the roughness-constrained sampling.

Ablation Studies:
Removing stage conditioning led to a collapse in transition success (dominated by STF). Removing roughness constraints maintained stage transitions but caused a significant drop in subtask success due to parameter mismatches (dominated by PPF). Removing feasibility constraints primarily increased contact instability. This confirms that the full SRDP architecture is necessary to suppress both failure modes simultaneously.

5. Significance and Claims

The paper asserts that SRDP represents a significant step forward in applying generative policies to complex, multi-stage manufacturing processes. Its significance lies in:

Bridging the Gap: It provides a unified solution for the co-occurring problems of latent stage uncertainty and unconstrained parameter drift, which have previously been addressed separately or not at all in high-dimensional generative policies.
Physical Credibility: By embedding roughness-driven constraints directly into the diffusion sampling process, the method ensures that generated actions are not only semantically correct for the task stage but also physically feasible and quality-compliant.
Generalizability: While demonstrated on polishing, the authors suggest the formulation serves as a general mechanism for incorporating stage-dependent objectives and physical constraints into generative robotic policies, potentially applicable to grinding, drilling, and deburring.

The authors conclude that SRDP improves task performance, parameter stability, and final surface quality, offering a robust framework for autonomous execution in high-precision manufacturing environments.