TransCurriculum

Multi-Dimensional Curriculum Learning for Fast & Stable Locomotion

Prakhar Mishra1,*, Amir Hossain Raj2, Xuesu Xiao2, Dinesh Manocha1
1University of Maryland, College Park
2George Mason University
IROS 2026
Paper Code

Presentation Video

Overview

TransCurriculum is a transformer-based, multidimensional curriculum-learning approach for agile quadrupedal locomotion. It adapts task sampling across velocity commands, terrain difficulty, and domain-randomization parameters using locally retrieved training history. We evaluate the approach on the Unitree Go1 in Isaac Gym and through zero-shot hardware deployment.

6.3 ± 0.2 m/s
Maximum simulated velocity
4.1 ± 0.05 m/s
Zero-shot hardware velocity
27% → 18%
Sim-to-real transfer loss

Approach

TransCurriculum uses a transformer-based teacher to model context–outcome history and adaptively sample task parameters across velocity commands, terrain difficulty, and domain randomization. The resulting curriculum trains a PPO locomotion policy for fast and stable sim-to-real deployment.

Overview of the TransCurriculum training pipeline
TransCurriculum training and task-sampling pipeline.

Zero-Shot Hardware Deployment

We deploy the learned policy zero-shot on the Unitree Go1 across rigid, deformable, inclined, and irregular terrain. The full multidimensional curriculum reaches 4.1 ± 0.05 m/s on hardware while improving stability and reducing sim-to-real transfer loss from 27% to 18%.

Unitree Go1 deployed across pebbles, a wooden slope, and rocky terrain
Zero-shot deployment across pebbles, an inclined wooden surface, and irregular rocky terrain.

BibTeX

@inproceedings{mishra2026transcurriculum,
  title     = {TransCurriculum: Multi-Dimensional Curriculum Learning for Fast and Stable Locomotion},
  author    = {Mishra, Prakhar and Raj, Amir Hossain and Xiao, Xuesu and Manocha, Dinesh},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026}
}