TAME-Nav

TAME-NavTemporal Asymmetric Mixture-of-Experts Distillation for Agile and Versatile Robot Navigation

Lingjie Zhang1,2, Jun Xue1,2, Shanze Wang2,3, Haoyu Xi1,2, Xinming Zhang2,4, Xiaoyu Shen2, Wei Zhang2,*

1 Shanghai Jiao Tong University 2 Eastern Institute of Technology, Ningbo 3 The Hong Kong Polytechnic University 4 University of Science and Technology of China
Paper Code Video

A single local-navigation policy for obstacle avoidance, narrow passages, and local-trap escape, using only onboard observations and a local goal.

Video

Download video

Method

Distill a privileged Teacher into a temporal Soft-MoE Student that can run on the robot.

Teacher–Student architecture: clean histories and causal obstacle tracks supervise a temporal Soft-MoE Student through latent distillation, with a shared Actor and a training-only Critic.
Teacher–Student framework. Click the figure to enlarge.

Privileged supervision

The Teacher combines clean observation histories with causal dynamic-obstacle tracks. The Student learns its navigation representation through latent distillation.

Temporal Soft-MoE

A GRU encodes ten observation frames. Eight softly routed experts combine complementary navigation features into one representation.

Diverse training scenes

600 procedural layouts cover six scene families. Asymmetric randomization adds sensing noise and delay to the Student while keeping Teacher supervision clean.

At deployment

Onboard observations + local goal → temporal Student → shared Actor + LiDAR CBF → velocity commands. No explicit object tracking or global-map input.

Six training scene families: dense clutter, dynamic crossings, concave traps, irregular narrow corridors, occluded interactions, and composite indoor layouts.
Six training scene families600 procedural layouts cover static constraints and dynamic interactions. Figure 2.

Citation

@misc{zhang2026tamenav,
  title  = {TAME-Nav: Temporal Asymmetric Mixture-of-Experts
            Distillation for Agile and Versatile Robot Navigation},
  author = {Zhang, Lingjie and Xue, Jun and Wang, Shanze and Xi, Haoyu
            and Zhang, Xinming and Shen, Xiaoyu and Zhang, Wei},
  year   = {2026},
  eprint = {2610.12877},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url    = {https://arxiv.org/abs/2610.12877}
}