Search papers, labs, and topics across Lattice.
This paper introduces ADAPT, an adaptive dual-projection architecture for humanoid locomotion that uses horizontal elevation maps and vertical distance maps to represent the environment. ADAPT learns to adjust its spatial sensing range, expanding it during fast motion and contracting it in cluttered environments. Results show ADAPT significantly reduces observation dimensionality and computational overhead, leading to faster training and robust zero-shot transfer to a Unitree G1 Humanoid compared to voxel-based baselines.
Humanoid robots can now nimbly navigate complex terrain with drastically reduced computational cost thanks to a novel adaptive sensing architecture.
Agile humanoid locomotion in complex 3D en- vironments requires balancing perceptual fidelity with com- putational efficiency, yet existing methods typically rely on rigid sensing configurations. We propose ADAPT (Adaptive dual-projection architecture for perceptive traversal), which represents the environment using a horizontal elevation map for terrain geometry and a vertical distance map for traversable- space constraints. ADAPT further treats its spatial sensing range as a learnable action, enabling the policy to expand its perceptual horizon during fast motion and contract it in cluttered scenes for finer local resolution. Compared with voxel-based baselines, ADAPT drastically reduces observation dimensionality and computational overhead while substantially accelerating training. Experimentally, it achieves successful zero-shot transfer to a Unitree G1 Humanoid and signifi- cantly outperforms fixed-range baselines, yielding highly robust traversal across diverse 3D environtmental challenges.
