Abstract
Imitation learning approaches achieve good generalization within the range of the training data, but tend to generate unpredictable motions when querying outside this range. We present a novel approach to imitation learning with enhanced extrapolation capabilities that exploits the so-called Equation Learner Network (EQLN). Unlike conventional approaches, EQLNs use supervised learning to fit a set of analytical expressions that allows them to extrapolate beyond the range of the training data. We augment the task demonstrations with a set of task-dependent parameters representing spatial properties of each motion and use them to train the EQLN. At run time, the features are used to query the Task-Parameterized Equation Learner Network (TP-EQLN) and generate the corresponding robot trajectory. The set of features encodes kinematic constraints of the task such as desired height or a final point to reach. We validate the results of our approach on manipulation tasks where it is important to preserve the shape of the motion in the extrapolation domain. Our approach is also compared with existing state-of-the-art approaches, in simulation and in real setups. The experimental results show that TP-EQLN can respect the constraints of the trajectory encoded in the feature parameters, even in the extrapolation domain, while preserving the overall shape of the trajectory provided in the demonstrations.
Introduction: The Extrapolation Challenge
In robot imitation learning, also known as Programming by Demonstration, a robot learns to replicate tasks from human examples. While state-of-the-art methods like Gaussian Mixture Models (GMMs) and standard Neural Networks (NNs) excel at interpolation—performing tasks within the range of demonstrated examples—they often fail during extrapolation.
When a robot is queried outside its training distribution (e.g., avoiding an obstacle taller than any seen during training), standard NNs often produce erratic or unnatural motions. This is largely because their activation functions, such as ReLU or Sigmoid, do not inherently capture the underlying physical logic of a trajectory. The Task-Parameterized Equation Learner Network (TP-EQLN) addresses this by fitting analytical mathematical expressions that generalize gracefully beyond training bounds.
Methodology: TP-EQLN Framework
The core innovation of this research is the integration of Equation Learner Networks (EQLNs) into the task-parameterized framework. Instead of standard weights, EQLNs utilize a library of analytical functions—including identity, sine, cosine, sigmoid, and multiplication—to discover the symbolic mathematical equation governing a motion.
Key Components
- Task Parameters ($\gamma$): The model accepts "features" of the task, such as obstacle height or goal position, alongside time ($t$).
- Three-Phase Training: The network undergoes a free convergence phase, followed by $L_1$ regularization (Lasso-like) to encourage sparsity, and finally a weight pruning phase to produce clean analytical expressions.
- Analytical Bias: By forcing the network toward trigonometric and sigmoidal functions, the model gains a "physical" bias. Robotic motions are often smooth or oscillatory, and these functions extrapolate far better than the piecewise-linear approximations of standard deep learning.
- Safety Constraints: A specialized "Penalty Epoch Term" in the loss function ensures predicted trajectories remain within the robot's physical workspace, even in the extrapolation domain.
Experimental Evaluation
The researchers evaluated TP-EQLN across six distinct tasks, ranging from synthetic 1D toy examples to complex real-world manipulation. These included simulated pick-and-place tasks with variable obstacle heights and goal positions, as well as opening boxes of varying sizes.
Real-world validation was conducted using a Franka Emika Panda robot. In one key experiment, the robot was tasked with pouring liquid into cups of varying sizes. This required the network to handle both position and orientation (quaternions) while generalizing to cup heights never encountered during the demonstration phase.
Key Results and Insights
TP-EQLN was compared against two industry standards: Task-Parameterized Gaussian Mixture Models (TP-GMM) and Conditional Neural Movement Primitives (CNMP).
- Superior Extrapolation: TP-EQLN demonstrated significantly lower Mean Squared Error (MSE) when tested outside the training range. In the pouring task, it maintained precise approach and pour points where CNMP produced distorted paths.
- Shape Preservation: Qualitatively, TP-EQLN trajectories remained "human-like" and smooth during extrapolation, preserving the intended motion profile.
- Automated Feature Mapping: Unlike TP-GMM, which requires manual definition of coordinate frames for every constraint, TP-EQLN automatically learns the relationship between high-level task parameters and the resulting trajectory.
Conclusion and Future Work
TP-EQLN represents a significant step toward robust robotic generalization. By leveraging analytical expressions, robots can safely handle novel environments while maintaining the integrity of learned skills. Future research will focus on incorporating dynamics and force control for contact-rich tasks, as well as adding uncertainty awareness to allow the robot to signal when it is unconfident in an extrapolated motion.
Summary updated