💻 computer science

Data and Learning Where it Matters for Contact-Rich Manipulation

This paper proposes a hybrid approach that combines traditional planning for free-space motion with automated offline deep reinforcement learning on a small, targeted dataset of critical contact-rich segments, achieving a 96% success rate across challenging real-world tasks while overcoming the brittleness and generalization issues of purely end-to-end learned policies.

Original authors: Oliver Hausdörfer, Linus Schwarz, Gabor Marko, Christian Dietz, Timo Class, Luka Hofer, Jim Yun-Jin Li, Johannes Hechtl, Ralf Römer, Angela P. Schoellig

Published 2026-07-20

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Oliver Hausdörfer, Linus Schwarz, Gabor Marko, Christian Dietz, Timo Class, Luka Hofer, Jim Yun-Jin Li, Johannes Hechtl, Ralf Römer, Angela P. Schoellig

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 world where robots are like clumsy toddlers trying to learn how to build a complex LEGO castle. They can easily pick up a brick and walk across the room (that's the easy part), but the moment they try to snap two pieces together, they often fumble, drop the pieces, or push too hard and break them. This is the heart of "contact-rich manipulation" in robotics: getting a robot to touch, push, and fit things together with the precision of a human hand. For years, scientists have tried to solve this by feeding robots massive amounts of video data, hoping the robot would learn by watching thousands of examples, much like a child learns by trial and error. But here's the catch: robots are expensive, time is money, and just throwing more data at the problem hasn't been working. The robots are still struggling with the tricky, high-precision moments, and they get confused when the scene changes even slightly.

This paper tackles that exact problem by asking a simple, almost obvious question: Why are we teaching the robot the easy parts the hard way? The authors suggest that instead of trying to learn the entire task from scratch using a giant, messy pile of data, we should split the job. We should use old-school, reliable math to handle the easy "walking around" parts, and only use the expensive, data-hungry learning for the tiny, critical moment when the robot actually has to push the pieces together. By focusing their learning efforts only where it truly matters, they found a way to make robots much more reliable without needing years of practice.

The "Focus on the Finish Line" Strategy

The researchers, working with teams from TU Munich and Siemens, discovered that most robot failures happen at one specific moment: the "critical segment." Think of it like a video game level where you can run freely through the open world, but the game ends if you miss a single, tiny jump. In robot tasks, the "free world" is moving an arm to grab an object, and the "tiny jump" is the moment of contact—like sliding a salt box into a tight shelf or snapping a LEGO brick onto a base.

The paper argues that current "end-to-end" learning methods (where the robot tries to learn the whole thing at once) are like trying to memorize the entire game map just to learn that one jump. It's inefficient and brittle. Instead, the authors propose a hybrid approach: use standard, pre-programmed planning for the easy movement, and only switch to a "learned" brain for the tricky contact part.

How they did it:

The Setup: They started with a single human demonstration of the task (like a teacher showing a student how to do it once).
The "Smart" Practice: Instead of having a human guide the robot every time, they let the robot replay that demo until it reached the tricky part. Then, the robot started "exploring" on its own. It tried a mix of random moves and smart guesses to figure out exactly how to push the object into place. This happened automatically, without a human holding the robot's hand.
The Learning: They used a technique called "offline Deep Reinforcement Learning." Imagine the robot watching a huge library of its own practice attempts (both the successes and the failures) and learning the best way to move after the data was collected, rather than learning while it was moving. This is safer and faster.
The Switch: When the robot is deployed, it uses a standard planner to grab the object. The moment it feels a "bump" (contact), it switches to its newly learned "expert brain" to finish the job. It knows when it's done by checking a "confidence score" (called a Q-function) to see if the job is successfully finished.

The Results: Speed, Precision, and Super-Reliability

The results were surprisingly effective. In just 2 to 2.5 hours of autonomous data collection (where the robot practiced on its own), the system achieved an average success rate of 96% across four difficult real-world tasks. These tasks included:

Shelf Stocking: Fitting a cardboard salt box into a tight, crowded shelf.
LEGO Stacking: Precisely placing a brick on top of another.
Fan Cover Assembly: Snapping a plastic cover onto a base, which involves bending and pressing deformable parts.

Compare this to the strongest existing methods (the "baselines"), which only managed a 55% success rate. The old methods often got the robot to the right spot but failed to push the object in all the way, or they broke the object by pushing too hard.

The authors also tested the robots in "out-of-distribution" scenarios—situations the robot had never seen before, like different background objects or slightly shifted positions. While the end-to-end learning robots got completely confused and failed, the authors' method kept its cool, maintaining high success rates. This suggests that by focusing on the specific mechanics of the contact, the robot learned a more robust skill that didn't rely on memorizing the whole scene.

Why This Matters

The paper explicitly argues against the idea that "more data" is always the answer. They show that simply scaling up data collection doesn't fix the problem if the data is scattered and unfocused. Instead, they prove that structure is key. By treating the easy parts of a task as "solved" and only applying heavy learning to the hard parts, they saved time and resources.

They also found that their method was much gentler on the objects. Because the robot learned the precise force needed for the contact, it applied 49% less force on average than the baseline methods. This is crucial for delicate items like cardboard boxes or plastic parts that can easily break if pushed too hard.

In short, this paper suggests that the future of robot learning isn't about building a bigger brain that knows everything; it's about building a smarter team where a reliable planner handles the walking, and a specialized, highly-trained expert handles the tricky handshake. It's a shift from "learning everything" to "learning what matters," and it seems to be a winning strategy for getting robots to do real-world jobs with human-like precision.

Technical Summary: Data and Learning Where it Matters for Contact-Rich Manipulation

Problem Statement

End-to-end robot learning, particularly when trained on large datasets, has shown promise in contact-rich manipulation tasks that are difficult to model analytically. However, these approaches often exhibit brittleness in high-precision scenarios and struggle with generalization to out-of-distribution (OOD) settings. The authors argue that these limitations stem not merely from a lack of data scale, but from a lack of structure in data collection. Specifically, end-to-end methods treat the entire task uniformly, whereas high-precision tasks typically fail only at the "critical segment" (e.g., tight clearances during insertion), while the remainder of the task involves unconstrained free-space motion that is easier to handle. Furthermore, current data collection pipelines often rely heavily on skilled human teleoperators, making them expensive and prone to variability.

Methodology

The paper proposes a compositional framework that explicitly exploits the structure of contact-rich tasks by separating them into two distinct phases: free-space motion and the critical contact-rich segment.

1. Task Decomposition and Planning

Free-Space Motion: For the unconstrained parts of the task (approach, grasp, transport), the system relies on off-the-shelf pose estimation (using language-guided segmentation and 6D pose estimation) and traditional motion planning. No robot-specific training data is required for these phases.
Critical Segment: The learned policy is applied only to the critical segment where contact occurs and precision is required.

2. Autonomous Data Collection

The authors introduce an automated data-acquisition scheme designed to collect dense data specifically at the critical segment without constant human intervention:

Initialization: A single demonstration is recorded via kinesthetic teaching to define the task and the scene layout.
Replay and Exploration: The system replays the demonstration until the critical segment is reached (defined by pose estimation uncertainty). It then executes a mixed greedy-random exploration policy ( $\pi_{explore}$ $π_{e x pl or e}$ ).
- With probability $p$ , the robot takes a random action.
- With probability $1-p$ , it takes a greedy action minimizing the distance to the goal state.
- A safety filter constrains the end-effector to a safe sphere to prevent drifting outside the critical segment.
Control: During these critical interactions, the system employs hybrid force-position control to ensure compliant behavior.
Reward: A sparse reward is used, triggered only upon task completion (distance to final state), eliminating the need for pre-trained vision-based reward classifiers.

3. Policy Learning

Algorithm: The system uses Offline Deep Reinforcement Learning (DRL), specifically building on Soft Actor-Critic (SAC) with improvements for sample efficiency (Layer Normalization in the critic and Randomized Ensembled Double Q-Learning).
Training: The policy is trained offline on the collected dataset. Since the setting is offline, the entropy coefficient is fixed rather than adapted online.
Deployment Switching: During execution, the system switches between the motion planner and the learned policy based on contact events (force thresholds) and Q-function values. A learned Q-function also serves as a success classifier to terminate the policy once the task is completed.

Key Contributions

Compositional Framework: A method that leverages dense data collection and offline DRL specifically for the critical task segment while relying on traditional motion planning for free-space motion.
Autonomous Data Collection Scheme: An automated pipeline that uses mixed greedy-random exploration to generate dense data at the critical segment, reducing reliance on skilled teleoperators.
Integrated Deployment: A sequential execution strategy that combines planning and learning using contact events and Q-function thresholds for policy switching, achieving high success rates across diverse tasks.

Experimental Results

The approach was evaluated on four challenging real-world tasks: Shelf Stocking, Lego Stacking, Fan Cover Assembly, and a difficult variant of Fan Cover Assembly involving scene distractors.

Success Rates: The proposed method achieved an average success rate of 96% across all tasks, with individual task rates ranging from 94% to 98%. This significantly outperformed the strongest baselines, which achieved a maximum average of 55%.
Data Efficiency: The method required only 2 to 2.5 hours of autonomous wall-clock time for data collection. In contrast, baselines like HIL-SERL required longer times and more human intervention.
Generalization (OOD): The method maintained high success rates in OOD scenarios (e.g., different object placements, distractors), whereas end-to-end baselines (Imitation Learning and DRL) struggled significantly, often dropping to near-zero success in difficult OOD conditions.
Force Application: The modular approach applied significantly less force to objects (average reduction of 49% or -4.7 N) compared to baselines, which is critical for handling deformable components like cardboard and PBT/PC molds.
Ablation Studies: Experiments confirmed that dense data coverage of the critical segment is essential. Pure random exploration or pure imitation (low exploration) failed to yield robust policies; a balance (random action probability $p \approx 0.8$ ) was optimal.

Significance and Claims

The paper claims that the gap in performance for high-precision contact-rich tasks is not solely a matter of data scale but of data structure and focus. By targeting data collection only where it matters (the critical segment) and leveraging traditional planning for the rest, the authors demonstrate that:

High success rates can be achieved with minimal autonomous data collection time (2–2.5 hours).
Reliance on skilled human teleoperators can be eliminated for the data collection phase.
Robustness to OOD scenarios is improved by narrowing the scope of the learned policy to local, contact-rich interactions rather than attempting to learn the entire task end-to-end.
This approach paves the way for targeted data collection strategies that make precision manipulation applications more feasible and robust.

The authors note limitations regarding tasks where sequential boundaries are hard to define (e.g., folding) and the dependency on the accuracy of off-the-shelf pose estimation for exotic or occluded objects. However, they posit that these pose estimation models can be fine-tuned more easily than training full robot policies.