Beyond Predefined Actions: Integrating Behavior Trees and Dynamic Movement Primitives for Robot Learning from Demonstration

TL;DR

The approach readily affords both learning to replicate low-level motions and combining partial demonstrations into a coherent and easy-to-modify overall policy, which enhances policy interpretability, modularity, and adaptability for autonomous systems.

Abstract

Interpretable policy representations like Behavior Trees (BTs) and Dynamic Motion Primitives (DMPs) enable robot skill transfer from human demonstrations, but each faces limitations: BTs require expert-crafted low-level actions, while DMPs lack high-level task logic. We address these limitations by integrating DMP controllers into a BT framework, jointly learning the BT structure and DMP actions from single demonstrations, thereby removing the need for predefined actions. Additionally, by combining BT decision logic with DMP motion generation, our method enhances policy interpretability, modularity, and adaptability for autonomous systems. Our approach readily affords both learning to replicate low-level motions and combining partial demonstrations into a coherent and easy-to-modify overall policy.

Figures (6)

• Figure 1: Experimental setup for evaluation: a Franka Emika Panda robotic manipulator equipped with a shovel, scoops a tennis ball and deposits it into one of two stations.
• Figure 2: Behavior Tree nodes: \ref{['fig:control-nodes']} Control nodes (Sequence, Parallel, Fallback) and \ref{['fig:execution-nodes']} Execution nodes (Condition, Action).
• Figure 3: Proposed method flowchart. (1) Data collection gathers demonstrations $T$, linking trajectories $t_i$ with state variables $c_i$. (2) DMP Learning recursively fits segmented trajectories with DMPs $\pi^{dmp}_{i,j}(\theta_{i,j})$, evaluated via DTW. (3) BT Learning trains a Decision Tree with CART and converts it to a BT policy $\pi^{bt}$ using RE:BT-Espresso.
• Figure 4: Recursive segmentation of $t_i$: split into $t_{i,j}$ if $DTW(t_{i,j}, \tau_{i,j}) > \epsilon$. Valid segments (green) link to DMPs when $DTW \leq \epsilon$, continuing on $t_{i,j+1}$.
• Figure 5: (a) Experimental task: The robot scoops a ball from S0, and transports it to S1 or S2. Trajectories: O1 (red), O2 (blue), and O3 (green) with an obstacle near S1. (b) Snapshot from a video demonstration showing O3.