ADAPT: Adaptive Dual-projection Architecture for Perceptive Traversal

Shuo Shao; Tianchen Huang; Wei Gao; Shiwu Zhang

cs.ROMar 17, 2026

ADAPT: Adaptive Dual-projection Architecture for Perceptive Traversal

Authors: Shuo Shao, Tianchen Huang, Wei Gao, Shiwu Zhang

Organizations: Department of Automation, University of Science and Technology of China, Anhui 230026, China. · Institute of Humanoid Robots, Department of Precision Machinery and Precision Instrumentation, University of Science and Technology of China, Hefei, Anhui 230026, China.

Abstract

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.

Figures & tables

Method	Perceptual Representation	Native Sampling	Sensory Dimension	Overhead	Adaptive Sensing
PIM [ 10 ]	Elevation Map	0.10 m	96	✗	✗
Beamdojo [ 20 ]	Elevation Map	0.10 m	225	✗	✗
Vb-com [ 16 ]	Elevation Map	0.10 m	225	✗	✗
Parkour [ 31 ]	Depth Image	48 $\times$ 64 px	3072	✓	✗
Omni [ 23 ]	Point Cloud	0.05 $\pi$ rad	251	✓	✗
Gallant [ 2 ]	Voxel Grid	0.05 m	40960	✓	✗

TABLE I: ADAPT comparison with existing works

Fig. 1: The proposed training architecture separates exteroceptive and proprioceptive inputs. Two MLP encoders process the perception maps independently. Their features are then fused with the proprioceptive state and passed to a GRU-based recurrent actor-critic. The actor predicts both the joint targets and the perception radius for the next step. The critic uses privileged, noise-free simulator information to produce a more accurate value estimate and stabilize training.

Fig. 2: Illustration of the adaptive perception radius mechanism. The vertical distance map is constructed on a cylindrical sector with a fixed number of cells ( $17\times 13$ ). A smaller radius (blue) provides higher spatial resolution for precise near-field sensing, while a larger radius (purple) extends the perception range at the cost of reduced resolution. The policy learns to dynamically adjust this radius based on the robot’s locomotion state.

Fig. 3: Examples of procedurally generated obstacles in our BarrierTrack training environment. The curriculum includes a diverse set of challenges such as stairs, gaps, hurdles, and constrained passages like beams and narrow gates, promoting the development of a versatile locomotion policy.

Terrain	Param. (m)	$\mathbf{p}_{\tau}^{\min}$	$\mathbf{p}_{\tau}^{\max}$
Stairs Up	Step Height $\uparrow$	0.10	0.30
Stairs Up	Step Length $\downarrow$	0.30	0.50
Drop	Drop Height $\uparrow$	0.10	0.60
Drop	Platform Length $\downarrow$	0.30	0.50
Jump	Wall Height $\uparrow$	0.05	0.40
Jump	Wall Depth $\uparrow$	0.10	0.30

TABLE II: Parameters for generating curriculum training terrains.

Method	Plane		Stairs		Drop		Jump
Method	$R_{\mathrm{succ}}$	$R_{\mathrm{trav}}$	$R_{\mathrm{succ}}$	$R_{\mathrm{trav}}$	$R_{\mathrm{succ}}$	$R_{\mathrm{trav}}$	$R_{\mathrm{succ}}$	$R_{\mathrm{trav}}$
PIM [ 10 ]	100 ( $\pm 0.0$ )	100 ( $\pm 0.0$ )	65.33 ( $\pm 37.57$ )	70.00 ( $\pm 34.61$ )	99.33 ( $\pm 3.59$ )	99.77 ( $\pm 1.26$ )	90.67 ( $\pm 16.92$ )	91.67 ( $\pm 14.62$ )
Parkour [ 31 ]	100 ( $\pm 0.0$ )	100 ( $\pm 0.0$ )	28.00 ( $\pm 28.10$ )	37.20 ( $\pm 24.93$ )	91.33 ( $\pm 14.31$ )	92.40 ( $\pm 13.52$ )	44.67 ( $\pm 20.45$ )	51.40 ( $\pm 17.49$ )
TSDF	100 ( $\pm 0.0$ )	100 ( $\pm 0.0$ )	40.00 ( $\pm 20.00$ )	47.43 ( $\pm 17.50$ )	56.67 ( $\pm 24.81$ )	63.33 ( $\pm 22.75$ )	43.33 ( $\pm 13.74$ )	48.63 ( $\pm 13.65$ )
Gallant [ 2 ]	100 ( $\pm 0.0$ )	100 ( $\pm 0.0$ )	70.00 ( $\pm 22.95$ )	76.30 ( $\pm 21.75$ )	80.00 ( $\pm 31.41$ )	83.23 ( $\pm 27.62$ )	68.00 ( $\pm 13.27$ )	74.10 ( $\pm 11.34$ )
Ours	100 ( $\pm 0.0$ )	100 ( $\pm 0.0$ )	94.67 ( $\pm 13.60$ )	95.90 ( $\pm 10.63$ )	100.00 ( $\pm 0.00$ )	100.00 ( $\pm 0.00$ )	98.67 ( $\pm 4.99$ )	98.90 ( $\pm 4.12$ )

TABLE III: Benchmarked Comparison in Simulation.

R_{\mathrm{succ}}

: success rate (%,

\uparrow

R_{\mathrm{trav}}

: normalized traversal progress (%,

\uparrow

Method	Train/iter (s, $\downarrow$ )	GPU Peak (MB, $\downarrow$ )
Parkour [ 31 ]	97.8	28791
Gallant [ 2 ]	34.6	36054
TSDF	246	38473
Ours	11.3	15870

TABLE IV: Training Efficiency Comparison. Train/iter: wall-clock time per training iteration (s,

\downarrow

). GPU peak: peak GPU memory usage (MB,

\downarrow

Type	Config	$v{=}0.5$ m/s	$v{=}1.0$ m/s	$v{=}1.5$ m/s
Fixed Radius	$R=1$	51.33 ( $\pm 17.65$ )	51.33 ( $\pm 23.49$ )	24.67 ( $\pm 22.32$ )
	$R=3$	81.33 ( $\pm 21.25$ )	72.67 ( $\pm 15.90$ )	56.67 ( $\pm 29.70$ )
	$R=5$	64.00 ( $\pm 18.90$ )	64.67 ( $\pm 26.17$ )	53.33 ( $\pm 30.26$ )
w/o Adap. Reward	no $r_{\text{adap}}$	92.67 ( $\pm 10.93$ )	71.33 ( $\pm 25.66$ )	42.67 ( $\pm 29.09$ )
Adaptive Radius	Ours	94.67 ( $\pm 10.24$ )	96.00 ( $\pm 8.00$ )	84.00 ( $\pm 19.60$ )

TABLE V: Ablation on adaptive sensing radius.

R_{\mathrm{succ}}

: success rate (%,

\uparrow

). Values are averaged over Beam, Pole, and Narrow Gate terrains.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
Number of environments	4096
Number of steps per policy update	24
PPO epochs	5
Number of mini-batches	4
PPO clip parameter ( $\epsilon$ )	0.2
Value function clip parameter	0.2

Appendix

TABLE VI: PPO Hyperparameters

Component		Specification
Recurrent Policy	RNN Type	GRU
	RNN Hidden Dims	256
	Actor MLP Hidden Dims	[128, 64, 32]
	Critic MLP Hidden Dims	[128, 64, 32]
	Activation Function	ELU
Horizontal Map Encoder	Encoder Type	MLP

Appendix

TABLE VII: Policy Network Structure

Dynamics Randomization
Parameter	Range (Uniform)
Robot Base Mass (kg)	[-1.0, 3.0]
Motor Strength Scale	[0.8, 1.2]
Ground Friction Coefficient	[0.1, 1.25]
External Perturbations
Push Interval (s)	5

Appendix

TABLE VIII: Parameters and Ranges for Domain Randomization

Category	Reward Term	Weight
Task Fulfillment	Linear Velocity Tracking	2.0
Task Fulfillment	Angular Velocity Tracking	0.5
Locomotion Stability	Base Orientation	-1.0
	Base Angular Velocity XY	-0.3
	Hip Neutrality	-1.0
Behavior Regularization	Joint Velocity	-1e-3

Appendix

TABLE IX: Detailed Reward Component Weights

Humanoid Robot Locomotion

Humanoid robot locomotion studies how bipedal robots walk, run, recover from disturbances, and traverse varied terrain.

See all 136 papers

TAGA: Terrain-aware Active Gaze Learning for Generalizable Agile Humanoid Locomotion
Jun 4, 2026Peizhuo Li, Hongyi Li, Mingfeng Fan +9Active Perception Humanoid Robot Locomotion
MarmotLab, National University of Singapore · Center of X-Mechanics, Zhejiang University · South China University of Technology
Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids
Aug 30, 2026Efe Ongan, Chong Zhang, Boyang Sun +3Humanoid Robot Locomotion Robotic Perception
Robotic Systems Lab, ETH Zurich, Switzerland · ETH AI Center, ETH Zurich, Switzerland · Computer Vision and Geometry Group, ETH Zurich, Switzerland
PRIOR: Perceptive Learning for Humanoid Locomotion with Reference Gait Priors
Mar 19, 2026Chenxi Han, Shilu He, Yi Cheng +2Humanoid Robot Locomotion Robot Locomotion
Tsinghua University · ZERITH Robotics · Shanghai University

Jun 4, 2026J/K move · Enter open · S save

TAGA: Terrain-aware Active Gaze Learning for Generalizable Agile Humanoid Locomotion

Peizhuo Li, Hongyi Li, Mingfeng Fan, Fangzhou Xu+8

MarmotLab, National University of Singapore · Center of X-Mechanics, Zhejiang University · South China University of Technology

Active Perception Humanoid Robot Locomotion Robot Policy Learning Terrain-Aware Robot Locomotion RL for Legged Locomotion Active Vision

Open paper AskPDF arXiv

Agile humanoid locomotion across diverse challenging terrain demands both wide perceptual coverage and precise local geometry understanding. Motivated by the way humans selectively look at relevant terrain during locomotion, we introduce TAGA, a Terrain-aware Active Gaze learning framework for Attention-based humanoid control. By fusing vision, proprioception, and motion commands, our framework guides the model to learn anticipatory cues and actively attend to specific areas of the height scan, selectively using these informative regions for the downstream network. This adaptively increases the information density of observations under tight onboard computational constraints, thus enabling fine-grained perceptive locomotion over larger-scale terrains. We find that such gaze behaviors can naturally emerge through reinforcement learning alone, without requiring additional supervision or explicit guidance, significantly improve training efficiency. As a result, the trained policy demonstrates robust and generalizable locomotion in simulation and on hardware, including reliable terrain-aware foothold selection, elevated-platform traversal, competitive sparse-foothold traversal, and the largest reported real-world gap traversal distance of 1.2m among perceptive humanoid locomotion systems, while maintaining stability under severe perceptual disturbances and environmental interference.

Abstract

Humanoid Robot Locomotion

Explore similar work

TAGA: Terrain-aware Active Gaze Learning for Generalizable Agile Humanoid Locomotion

Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids

PRIOR: Perceptive Learning for Humanoid Locomotion with Reference Gait Priors