Abstract
Training robots to perform adaptable tasks usually requires dozens of demonstrations to cover various spatial scenarios. This paper introduces a novel method to learn Task-Parameterized Gaussian Mixture Models (TP-GMM) from as few as two demonstrations by augmenting the small dataset with synthetic movements. By "imagining" how a skill looks in new, valid configurations—a process called Random Frames (RF)—the robot significantly improves its ability to generalize to unseen situations, such as assisting a human with dressing.
Introduction
Learning from Demonstration (LfD) allows robots to acquire complex skills without manual programming. A popular framework for this is TP-GMM, which encodes motions relative to specific "reference frames" (e.g., the position of a cup or a person’s hand). This allows the robot to adapt its trajectory when those objects move.
However, TP-GMM is traditionally data-hungry. To learn a robust model, a human must demonstrate the task in many different spatial configurations. In real-world settings like factories or homes, collecting 20+ demonstrations is tedious and often impractical. Without sufficient data, the model fails to capture the statistical "essence" of the task, leading to poor performance in novel situations.
Methodology
The authors propose an iterative framework that grows a minimal initial dataset into a robust one using synthetic data generation. The process follows four key steps:
- Initial Training: The model is initialized with a very small set of expert demonstrations (e.g., only 2).
- Synthetic Data Generation: The paper evaluates three methods for creating training data: adding white noise, Random Frames (RF), and a combination of both. RF is the core innovation; the robot uses its current (imperfect) model to generate a trajectory in a new, randomly sampled valid situation (e.g., moving the target object to a new spot within reach).
- Selection Criteria: The robot retrains its model with the new synthetic data. If the new model reproduces the original expert demonstrations with lower error (measured via RMS or Dynamic Time Warping), the synthetic data is retained.
- Iteration: This process repeats until a maximum number of demonstrations or iterations is reached.
The RF method is particularly effective because it forces the model to reconcile perspectives from different reference frames, helping it distinguish which parts of the motion are fixed to specific environmental anchors.
Experimental Setup
The researchers evaluated the algorithm on two distinct benchmarks:
- Simulated 2D Task: A robot must move from a start point to a goal while navigating a U-shaped constraint. This used 6 expert demonstrations (3 for training, 3 for validation).
- Robotic Dressing Task: A physical robot arm assisting a mannequin in putting on a short-sleeve shirt. The task parameters were the positions of the mannequin's shoulder and wrist. This used only 2 initial expert demonstrations.
Results and Discussion
The Power of "Imagination"
While adding noise is a common machine learning technique, the authors found that in TP-GMM, simple noise injection can lead to overfitting, magnifying unwanted "wiggles" in the motion. In contrast, generating data in new random frames (RF) provides the structural variety needed for the model to realize, for example, that the beginning of a reach depends on the robot's base, while the end depends entirely on the target's location.
Performance Gains
- Generalization: In simulations, the RF and RF+Noise methods reduced validation error in new situations by roughly 13-14% compared to the initial model. Conversely, simple noise injection actually increased error by 18%.
- Success Rate: In the complex dressing task, a model trained on just 2 demonstrations had a success rate of only 20%. After augmenting the data using the proposed algorithm, the success rate jumped to 80%.
- Computational Efficiency: The algorithm runs in approximately 20 seconds on a standard PC, making it significantly faster than the time required for a human to perform extra manual demonstrations.
Conclusion
This research demonstrates that robots can learn robust, generalized skills from extremely limited data by leveraging synthetic data generation within a task-parameterized framework. While the current method requires a pre-defined task space and operates in an open-loop fashion, it paves the way for more efficient robot training in dynamic environments. Future work may explore automatically adjusting model complexity as more synthetic data is integrated.
Summary updated