ViBe: Visual Behavior Adaptation
for Perceptive Humanoid Whole-body Control

Lokesh Krishna, Sarvesh Venkatesan, An Zhang, Quan Nguyen
University of Southern California
Paper arXiv Live Demo Code (Coming Soon)

Overview Video

Interactive Demo: Repose Cube

Simple planner, learned controller: a fixed rule-based planner selects motions; ViBe handles perceptive whole-body control.

Approach

TL;DR
Post-training humanoid whole-body trackers for perceptive control.
Task-relevant feedback is learnt via cross-attention extractor from pre-trained vision encoders and grafted with low-rank adapters for parameter-efficient fine-tuning .

ViBe system diagram connecting task environments and observations to the visual extractor, task adapter, planner, and pretrained whole-body controller

Simulation Results

Learnt attention maps are shown as insets over their corresponding third-person rollouts.

Repose Cube
Parkour
Dodge Ball
Walk
Omni-Object Loco-Manipulation

Adapt to Environment Constraints

Parkour
Walk

Handle External Dynamics

Repose Cube
Pick-Place Suitcase
Dodge Ball
Pick-Place Trash Can

Visual Robustness

Parkour — Outdoor
Repose Cube — Disco Lighting
Walk — Outdoor
Repose Cube — Low Light

Repose Cube with a Planner

Get the Green Face on Top
Get the Yellow Face on Top
Get the Orange Face on Top
Get the Red Face on Top
Get the Pink Face on Top
Get the Blue Face on Top

Citation

@article{krishna2026vibe,
  title         = {ViBe: Visual Behavior Adaptation for Perceptive Humanoid Whole-Body Control},
  author        = {Krishna, Lokesh and Venkatesan, Sarvesh and Zhang, An and Nguyen, Quan},
  journal       = {arXiv preprint arXiv:2609.09918},
  year          = {2026},
  eprint        = {2609.09918},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}