Xidian University
B.Eng. · OMEGA Lab
National Scholarship
Personal introduction · 2026
Yue Su
Learning world models that guide action.
Experience
Across academia and industry,
one continuous research thread.
MVIG · SJTU
Research Assistant
Robot learning
Astribot
Research Intern
Whole-body manipulation
ByteDance Seed
Research Intern
Seed Robotics
MMLab · HKU
PhD
World models & action
MBA · 01 / 02
Motion before action.
Observation
Objects define the task
The same robot motion can be right or wrong depending on the desired object motion.
Condition
Make the future explicit
Generate a dense object-motion representation before synthesizing the action.
Thesis
Imagination can guide control
A predictive intermediate signal makes manipulation more purposeful.
MBA · 02 / 02
Two diffusion processes, one manipulation condition.
Step 1
Motion diffusion
Predict object motion from the current observation and language instruction.
Step 2
Action diffusion
Condition the manipulation policy on the generated motion representation.
Property
Plug and play
Attach MBA to existing manipulation policies without redesigning the action backbone.
Dense Policy · 01 / 02
Action does not have to be generated left-to-right.
Problem
Long trajectories are hard
Small early errors compound when every action depends only on its predecessor.
Insight
Keyframes carry structure
Begin with sparse, informative moments that anchor the overall manipulation.
Reframe
Order is a design choice
Autoregressive generation can follow task structure rather than time alone.
Dense Policy · 02 / 02
Expand from sparse keyframes—in both directions.
Bidirectional generation
Sparse → dense.
Predict key actions first, then recursively fill the intervals to form a coherent trajectory.
ICCV 2025 paper ↗Real-robot behavior
Coherent manipulation.
The policy preserves global task structure while refining detailed local actions.
See all demos ↗DSPv2 · 01 / 02
Whole-body mobile manipulation changes the problem.
Perception
See what matters
Mobile settings introduce larger visual variation and longer interaction horizons.
Coordination
Move as one system
Base, torso, and arm actions must form a coherent whole-body trajectory.
Generalization
Leave the training setup
A useful policy must transfer across objects, tasks, and environments.
World Guidance · One view
World modeling inside the condition space.
Compact
Less is more
Compress future observations into a non-redundant predictive space.
Predictable
Action-centric latent
The VLA predicts conditions that are intrinsically relevant to action.
Fine-grained
Guidance that acts
Future-aware conditions support precise and generalizable manipulation.
World Guidance · Method
Condition in. Condition out.
Complete inference.
Stage I
Condition in
Inject future observations into the action inference pipeline.
Stage II
Condition out
Predict compressed future conditions alongside actions.
Transfer
Guidance retained
Carry future knowledge into the VLA without requiring future frames at test time.