Overview
Continual robotic manipulation learning
The project studies how VLA agents acquire a sequence of tasks while preserving useful prior knowledge.
Vision-Language-Action
Stellar VLA.
Continually evolving skill knowledge for robot manipulation, with task-centric and task-skill variants for stable lifelong VLA learning.
01
Motivation
Why continual VLA needs reusable task-skill knowledge rather than growing task modules.
02
Method
T-Stellar and TS-Stellar model task knowledge and task-skill structure for adaptive action routing.
03
Experiments
Dual-arm real-robot tasks and LIBERO continual imitation learning evaluations.
Motivation
Lifelong robot policies need stable skill memory.
Vision-language-action models can inherit broad manipulation knowledge from pretraining, but efficient continual learning remains difficult. When new tasks arrive sequentially, the robot must adapt without overwriting old behaviours or adding a separate module for every task.
Stellar VLA reframes this as knowledge modelling. Instead of treating each task as isolated, the policy learns a compact knowledge space that captures relationships between tasks and skills. That knowledge then guides expert routing for action prediction.
Real robot evidence
Knowledge retention across tasks
Real-world curves compare continual learning behaviour across seven dual-arm manipulation tasks.
Method
Task knowledge, skill structure, and knowledge-guided routing.
Architecture
Stellar VLA pipeline
Language and visual observations are encoded into task-centric representations, then aligned with a learned knowledge space for action prediction.
Knowledge space
Task and subskill organization
The learned representation exposes task-level clusters and shared subskill relationships in long-horizon manipulation sequences.
T-Stellar
Task-centric modelling
The flat variant learns task-relevant knowledge that helps specialize action prediction without expanding the policy for each new task.
TS-Stellar
Task-skill hierarchy
The hierarchical variant models how tasks share reusable subskills, which is important for long-horizon manipulation.
Expert routing
Knowledge-guided action head
Semantic embeddings and knowledge relationships guide which motion experts should be emphasized for a given task.
Real-World Experiments
Seven dual-arm manipulation tasks after continual learning.
Task 01
Pick up Stick
Grasping a stick with the dual-arm robot setup.
Task 02
Handover Stick
Coordinated handover behaviour between robot arms.
Task 03
Pick up Bag
Generalization to deformable or flexible objects.
Task 04
Handover Toy
Handover with object and scene variation.
Task 05
Pick Toy Place Plate
Pick-and-place with coordinated placement constraints.
Task 06
Pick Both Object
Dual-arm coordination with multiple objects.
Task 07
Pull Stick from Bag
Contact-rich extraction after learning previous tasks.
Results
Real-world and simulation evaluations.
Real world
Success-rate comparison
The real-robot evaluation compares continual manipulation performance across baseline policies and Stellar variants.
Simulation
LIBERO continual learning
Simulation studies cover goal, long-horizon, and multi-task LIBERO settings with limited replay.