Unlock Reliable Skill Inference for Quadruped Adaptive Behavior by Skill Graph

TL;DR

This work introduces the Robot Skill Graph (RSG), a knowledge-graph-inspired framework that organizes a large repertoire of quadruped robot skills and enables reliable inference and fast adaptation to new tasks and environments. By combining environment/task representations with skill policies through TransH-based relations and a contrastive training objective, RSG supports autonomous skill discovery, robust execution, and rapid learning via Bayesian optimization or PPO fine-tuning. The approach yields a broad, semantically structured skill set (320 skills across 12 environments and 31 tasks) and demonstrates effective parkour-style locomotion, robust behavior under disturbances, and quick adaptation to unseen scenarios. The work advances autonomous robot reasoning and learning by marrying domain knowledge with graph-based representations for scalable, context-aware skill reuse and online skill synthesis.

Abstract

Developing robotic intelligent systems that can adapt quickly to unseen wild situations is one of the critical challenges in pursuing autonomous robotics. Although some impressive progress has been made in walking stability and skill learning in the field of legged robots, their ability for fast adaptation is still inferior to that of animals in nature. Animals are born with a massive set of skills needed to survive, and can quickly acquire new ones, by composing fundamental skills with limited experience. Inspired by this, we propose a novel framework, named Robot Skill Graph (RSG) for organizing a massive set of fundamental skills of robots and dexterously reusing them for fast adaptation. Bearing a structure similar to the Knowledge Graph (KG), RSG is composed of massive dynamic behavioral skills instead of static knowledge in KG and enables discovering implicit relations that exist in between the learning context and acquired skills of robots, serving as a starting point for understanding subtle patterns existing in robots' skill learning. Extensive experimental results demonstrate that RSG can provide reliable skill inference upon new tasks and environments, and enable quadruped robots to adapt to new scenarios and quickly learn new skills.

Figures (12)

• Figure 1: Overall framework for the construction and application of RSG.a, Collect a rich and diverse set of fundamental skills through the DRL approach with each skill consisting of a task, an environment and a policy network. b, Different skills are connected through relationships between environmental entities and task entities. c, Illustration of the RSG structure. d, Given a new task and environment query, the RSG calculates the match between existing fundamental skills and new required skills. e, Finally, these skills inferred by RSG will be executed, composited, or fine-tuned respectively according to the matching degree. The newly learned skills will be also added to RSG for future usage.
• Figure 2: Overview of the presented RSG.a, Fundamental skills are divided into locomotion skill types (walking, running, etc.) and flexible skill types (rolling, posture recovery, etc.). The context-aided estimator network (CEN) 25_nahrendra2023dreamwaq is used to enhance the environmental representation. We also utilize adversarial motion priors (AMP) 50_escontrela2022adversarial to provide expert-style behavior for locomotion skill tasks. b, To construct the RSG, the environment is described as friction, flatness, and slope, and the task is described as eleven consecutive CoM trajectory points. The environment and task descriptions are then mapped into latent variables respectively (with 48 dim. here as a hyperparameter). The TransH is leveraged to complete the relationship construction between skill entities, task entities, and environment entities. c, For a new skill query, we use a score function to measure the match between the required new skill and the existing fundamental skills (i.e., skill inference score). For the application, according to the high, medium, and low scores of the inferred skills, the action modes adopt the methods of skill execution, BO composition, and RL finetuning respectively.
• Figure 3: Visual analysis of RSG representation and score assignment.a-i, T-SNE visualization of environment, skill, and task representations in raw form, RSG (Ours, w/TransH), and RSG w/TransE, respectively. j-l, score assignments by SBM, RSG (Ours), and RSG w/TransE for backward left walking on grassland across task queries (31 task categories in total), respectively. m-o, score assignments by SBM, RSG (Ours), and RSG w/TransE for forward walking on downstairs across environment queries (12 environment categories in total), respectively. The box represents 25% quantile, median, 75% quantile, and whiskers are data within 1.5 times of the Inter-Quantile Range. The arrows after each row mark the corresponding legend. Each type is highlighted by the background color of the box plot.
• Figure 4: Broad distribution of fundamental skills in RSG. Robot tasks include rolling, standing, walking, turning, small steps, etc. The environment includes unstable debris, narrow passages, single-plank bridges, sponge mats, dirt piles, lawns, mud, pebbles, stairs, soft clays, etc.
• Figure 5: The deployment of RSG for robot parkour.a, Autonomous inference of real-world robot skills. Two inferred skills and the corresponding first-person images are shown. b, Robust execution of diverse robot skills in a realistic parkour task. c, The profiles of robot body velocity and right front leg torque corresponding to b. d, Average of foot position (right front leg) trajectories corresponding to several skills in b. e-g, Ten snapshots of robot motion in simulation, CoM position trajectories, and t-SNE visualizations of the executed skills.