💻 computer science

COMPASS: Confined-space Manipulation Planning with Active Sensing Strategy

This paper introduces COMPASS, a multi-stage framework that integrates near-field collision mapping, utility-driven active sensing, and constrained manipulation optimization to significantly improve success rates in robot manipulation tasks within confined and cluttered environments.

Original authors: Qixuan Li, Chen Le, Dongyue Huang, Jincheng Yu, Xinlei Chen

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qixuan Li, Chen Le, Dongyue Huang, Jincheng Yu, Xinlei Chen

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 a robot arm trying to grab a toy that is buried deep inside a messy, narrow box. The box is so full of other objects that the robot can't see the toy, and its own arm is so long and bulky that it might bump into the sides of the box if it moves carelessly. This is the "confined space" problem that the paper COMPASS tries to solve.

Here is how the paper explains their solution, broken down into simple concepts and analogies:

The Problem: The "Blindfolded Diver"

Most robots are trained in open rooms where they can see everything from a distance, like a person standing in a field looking at a tree. But in a tight, cluttered space, the robot is like a diver in a dark, narrow cave.

The Blind Spot: The robot can't see the target because other objects are blocking the view (Perception Occlusion).
The Tight Squeeze: The robot's arm is long and clumsy in a small space. If it tries to reach for the target without checking first, it might crash into the walls (Kinematic Constraints).

Old methods either tried to guess the path without looking (which leads to crashes) or just looked for the most "interesting" new spot without caring if the robot could actually reach it (which leads to getting stuck).

The Solution: COMPASS (The Careful Explorer)

The authors created a three-step system called COMPASS to help the robot act like a smart, cautious explorer rather than a reckless one.

Step 1: The "Feeling Around" Scan (Near-Field Awareness)

Before the robot tries to move far, it does a quick, safe "wiggle" with just its wrist (the hand part of the arm) while keeping its body still.

Analogy: Imagine you are in a pitch-black room. Before you take a big step, you wave your hands in front of you to feel where the walls are.
What it does: This builds a tiny, safe map of the immediate area right next to the robot. It ensures the robot doesn't crash into something right next to it before it even starts its main journey.

Step 2: The "Smart Detective" Search (Active Sensing)

Now that the robot knows the immediate area is safe, it starts looking for the target. But it doesn't just pick random spots to look at. It uses a special planner called MUE-RRT.

Analogy: Think of a detective looking for a lost key in a messy room. A bad detective just spins around randomly. A smart detective asks: "If I move my head here, will I see a new corner? Is it easy for my neck to turn that way? And will this angle help me grab the key once I find it?"
What it does: The robot picks the next best spot to look at based on three things:
1. Information: Will I see new things?
2. Safety: Can my arm actually get there without hitting anything?
3. Readiness: Is this a good position to grab the object later?
  It moves smoothly from one spot to another, building a 3D map as it goes, until it spots the target.

Step 3: The "Perfect Grip" (Constrained Manipulation)

Once the robot finds the target, it doesn't just lunge for it. It calculates a very specific way to grab it.

Analogy: Imagine trying to grab a coin from a jar full of marbles. You can't just shove your whole hand in; you have to angle your fingers perfectly so they don't get stuck on the other marbles.
What it does: The system checks the map it just built and calculates a "safe path" for the gripper. It makes sure the robot's arm can physically reach the target without hitting the obstacles it just mapped out.

The Proof: The "Obstacle Course"

To prove their system works, the authors didn't just test it once. They built a video game-style benchmark with four levels of difficulty:

Level 1: A little bit of clutter.
Level 2: Very hidden targets (hard to see).
Level 3: Very tight spaces (hard to move).
Level 4: The "Nightmare Mode" (hard to see AND hard to move).

They tested their robot in a computer simulation and then in the real world with a real robot arm.

The Results

Success Rate: Their method was much better than other methods. In the simulations, they succeeded 24.25% more often than the next best method.
Speed: They found the target faster and moved less distance to do it.
Real World: When they tried it on a real robot in a messy box, it succeeded 80% of the time.

Summary

The paper argues that to grab things in tight, messy spaces, a robot needs to stop thinking of "seeing" and "moving" as two separate jobs. Instead, it must be a unified team: it must look for places that are safe to move to, move in a way that helps it see better, and only grab the object when it is absolutely sure it can do so without crashing. COMPASS is the brain that coordinates this careful dance.

Technical Summary: COMPASS

Problem Statement

Robotic manipulation in confined and cluttered environments presents a dual challenge: partial observability and complex kinematic constraints. In such spaces, surrounding obstacles often occlude the target, rendering methods that assume full observability ineffective. Simultaneously, the manipulator's own embodiment and the surrounding obstacles impose tight kinematic constraints, drastically limiting the reachable workspace. Existing paradigms fall short in addressing this dilemma:

End-to-end learning methods often rely on fixed third-person cameras (unavailable in confined spaces) and lack training data for complex obstacle interactions.
Planning-based methods typically assume complete world models and known target poses, ignoring the necessity of active exploration to discover the goal.

There is a critical need for an integrated perception-motion-manipulation approach that allows a manipulator to incrementally explore an unknown, confined space, safely build a scene understanding, and execute constrained manipulation.

Methodology: COMPASS Framework

The authors propose COMPASS (Confined-space Manipulation Planning with Active Sensing Strategy), a multi-stage framework designed to bridge the gap between exploration and manipulation. The system operates through three sequential stages:

1. Near-Field Awareness Scan

To mitigate collision risks in an initially unknown environment, the framework begins with a safety-oriented scan.

Objective: Build a local collision map by finding a minimal set of joint configurations that maximize volumetric coverage of a predefined safety envelope around the robot's base.
Mechanism: The robot's body configuration (base and shoulder) is held constant, while the wrist configuration space is sampled. A greedy optimization selects wrist waypoints that cover the largest remaining unknown volume within the safety envelope. This ensures the robot can move safely before attempting global exploration.

2. Manipulation-Utility Exploration RRT (MUE-RRT)

Once a local map is established, the system employs a sampling-based planner to find the target.

Frontier Detection: The system identifies frontiers (boundaries between known free space and unknown space) in an OctoMap constructed from wrist-mounted depth images.
Tree Expansion: A Rapidly-exploring Random Tree (RRT) is dynamically built, biased towards frontier regions and task priors.
Multi-Objective Utility Function: The Next-Best-View (NBV) is selected by maximizing a utility function $U(x)$ $U (x)$ that balances:
- Information Gain ( $G$ ): Volume of new space observable.
- Exploration Momentum ( $D$ ): Encourages smooth, continuous paths.
- Manipulability ( $M$ ): Ensures the robot remains far from singularities (using the Yoshikawa index).
- Motion Cost ( $C$ ): Path length in joint space.
- Heuristic Guidance ( $H$ ): Biases exploration towards specific points of interest.
Asynchronous Detection: While the robot moves, an observation buffer stores synchronized image-pose pairs. A detection model (e.g., YOLO) processes this buffer to identify the target, ensuring transient views are not missed due to inference latency.

3. Constrained Manipulation Pose Generation

Upon target detection, the system generates a feasible grasp pose.

Candidate Generation: A grasp detection network (GraspNet) generates candidate grasps based on depth data.
Constrained Optimization: Each candidate is refined via a constrained optimization problem to ensure:
- Kinematic Feasibility: The inverse kinematics solution exists within the free configuration space ( $Q_{free}$ ).
- Approach Direction: The grasp approach vector aligns with the object's orientation.
- Collision Avoidance: The entire trajectory (via forward kinematics) remains within the free space of the current map ( $M_{free}$ ).

Key Contributions

COMPASS Framework: A novel integrated framework for exploration and manipulation in confined spaces, enabling manipulators with partial observation to safely explore unknown environments and perform constrained manipulation.
MUE-RRT Planner: A manipulation-centric exploration planner that generates smooth, safe trajectories by optimizing for a multi-objective utility function that balances information gain with kinematic feasibility and motion cost.
Principled Benchmark: A systematically designed benchmark for confined-space manipulation, featuring four levels of increasing difficulty that isolate and combine Perception Occlusion and Kinematic Constraints.

Experimental Results

The method was evaluated in simulation (Isaac Sim) and real-world experiments using a Franka Panda manipulator.

Simulation Results

Success Rate: Compared to baselines (Fixed-Views, Information-Gain only, and Geometric Sampling), COMPASS increased the manipulation success rate by 24.25% over the Geometric Sampling Exploration (GSE) baseline and 39.25% over the Information-Gain (IG) baseline.
Efficiency: The framework reduced the average Time-to-Find-Target (TFT) by 24.6 seconds compared to baselines.
Motion Performance: COMPASS maintained a higher Motion Planning Success Rate (MPSR) and Average Manipulability (AM), particularly in high-constraint scenarios (Level 4), where baselines frequently failed to find valid motions.

Real-World Results

In real-world confined environments (Level 4 scenarios), the system achieved an 80% overall success rate across 10 trials per scenario.
Failures were primarily attributed to sensor noise causing motion planning failures or convergence to local optima in the exploration policy.

Ablation Studies

Near-Field Awareness: Disabling this stage reduced the Motion Planning Success Rate from 61.2% to 54.7% and the Target Detection Success Rate from 94.1% to 68.3%, confirming its critical role in safety.
Constrained Grasp Optimization: Removing grasp constraints led to a sharp drop in grasp success rate (from 70% to 20%) and a significant increase in motion collisions and object drops.

Significance and Claims

The paper claims that COMPASS addresses a critical gap in robotic manipulation by shifting from simple "perception-then-motion" pipelines to a truly integrated perception-motion-manipulation approach. The authors assert that their method is the first to present such an integrated solution specifically for confined-space tasks. By systematically evaluating performance across a progressive benchmark, the work demonstrates that considering manipulation utility during the exploration phase is essential for success in environments characterized by severe occlusion and tight kinematic constraints. The framework proves capable of performing active sensing and manipulation in real-world confined environments where traditional methods fail.