AME-2: Agile and Generalized Legged Locomotion via Attention-Based Neural Map Encoding

TL;DR

AME-2 tackles the challenge of agile, generalized legged locomotion under occlusions and sparse footholds by marrying an attention-based map encoder with a lightweight, uncertainty-aware neural mapping pipeline, all trained within a teacher–student RL framework. The approach preserves modular perception and control while enabling sim-to-real transfer, demonstrated on both a quadruped (ANYmal-D) and a biped (TRON1). Key contributions include the AME-2 encoder design, a fast, uncertainty-aware mapping pipeline, and a training strategy that delivers deployable policies with strong generalization to unseen terrains. The results show state-of-the-art agility and robust generalization, with emergent behaviors like active perception and whole-body skills, highlighting practical impact for real-world legged robots across diverse environments.

Abstract

Achieving agile and generalized legged locomotion across terrains requires tight integration of perception and control, especially under occlusions and sparse footholds. Existing methods have demonstrated agility on parkour courses but often rely on end-to-end sensorimotor models with limited generalization and interpretability. By contrast, methods targeting generalized locomotion typically exhibit limited agility and struggle with visual occlusions. We introduce AME-2, a unified reinforcement learning (RL) framework for agile and generalized locomotion that incorporates a novel attention-based map encoder in the control policy. This encoder extracts local and global mapping features and uses attention mechanisms to focus on salient regions, producing an interpretable and generalized embedding for RL-based control. We further propose a learning-based mapping pipeline that provides fast, uncertainty-aware terrain representations robust to noise and occlusions, serving as policy inputs. It uses neural networks to convert depth observations into local elevations with uncertainties, and fuses them with odometry. The pipeline also integrates with parallel simulation so that we can train controllers with online mapping, aiding sim-to-real transfer. We validate AME-2 with the proposed mapping pipeline on a quadruped and a biped robot, and the resulting controllers demonstrate strong agility and generalization to unseen terrains in simulation and in real-world experiments.

Figures (20)

• Figure 1: Our method enables agile and generalized legged locomotion across diverse terrains with onboard sensing and computation.
• Figure 2: An overview of our system. We use RL to train a teacher policy with ground-truth mapping in simulation, and a student policy under the teacher's supervision with our proposed neural mapping. The policies output joint-level actions that actuate the robots to reach position and heading goals, as illustrated by the yellow wireframes.
• Figure 3: An illustration of our AME-2 policy architecture. Left top: The general abstract of our policies. Proprioceptive observations are encoded into a proprioception embedding, which is used together with the mapping as the inputs of the AME-2 encoder to produce the map embedding. The proprioception embedding and the map embedding are then concatenated and fed into an MLP to generate actions. Left bottom: The AME-2 encoder design. We extract both local features and global features from the map, and then use the global features and the proprioception embedding to produce the attention-weighted local features. $L$ and $W$ are the length and width of the map, $d_{map}$ is the dimension of map representation (3 for the teacher, 4 for the student), $d_{PE}$ is the dimension of proprioception embedding. Right: The proprioception encoder designs. We have different designs for teacher and student policies to facilitate sim-to-real transfer, while both designs can fit into our overall architecture.
• Figure 4: Training and Test Terrains. We train our locomotion policies on primitive terrains of three categories: 1 dense, 2 climbing, and 3 sparse. To test generalization, we evaluate the policies on four terrains that are either entirely unseen or combinations of training and/or unseen terrains.
• Figure 5: The controller can reuse the built map for the same terrain, enabled by our mapping randomization. (a) When the robot moves in the direction of the dotted orange arrow, it observes a partial map of the staircase (visualized on the right). (b) After going down along the solid orange arrow, it perceives a complete map of the staircase (visualized on the right) when attempting to move back along the dotted orange arrow.