aiqbench analysis
ManiSkillFormer advances robotic manipulation by replacing end-to-end learning with a neuro-symbolic approach that uses LLMs to generate geometric contracts for reusable motion templates. This architecture enables zero-shot performance on diverse tasks like pick-and-place and functional manipulation without requiring per-object demonstrations or fine-tuning. The shift toward composable skills via explicit perception-action interfaces represents a significant step toward generalizable robotics grounded in AI reasoning.
ManiSkillFormer: Demonstration-Free Compositional Manipulation via Task-Conditioned Geometric Contracts
arXiv:2609.16331v1 Announce Type: new Abstract: We present ManiSkillFormer, a neuro-symbolic framework for demonstration-free and compositional robotic manipulation. Instead of learning end-to-end visuomotor policies, ManiSkillFormer introduces task-conditioned geometric contracts that explicitly structure the interface between perception and action. Each manipulation skill declares the semantic geometric primitives required for execution, such as object keypoints and surface normals. Building on human-defined skill structures, LLM agents generate these contracts and corresponding motion templates for different objects and task contexts. These contracts guide the perception module to ground task-relevant 3D primitives from observations, which are then used to instantiate reusable motion templates stored in a skill library. We evaluate ManiSkillFormer on Galaxea R1-Lite dual-arm robot across three settings: zero-shot pick-and-place over 8 object categories with 30 different instances, functional manipulation tasks including unscrewing, pouring, pressing, and folding, and 3 long-horizon tasks. ManiSkillFormer achieves higher average success rates than the evaluated baselines and two ablated pipelines: 88.24% for demonstration-free pick-and-place, 75.00% average success on functional manipulation and 50--80% completion rates across the long-horizon tasks. These results show that our design enables composable and reusable manipulation across objects and tasks without per-object policy fine-tuning or additional robot demonstrations.
Abstract:We present ManiSkillFormer, a neuro-symbolic framework for demonstration-free and compositional robotic manipulation. Instead of learning end-to-end visuomotor policies, ManiSkillFormer introduces task-conditioned geometric contracts that explicitly structure the interface between perception and action. Each manipulation skill declares the semantic geometric primitives required for execution, such as object keypoints and surface normals. Building on human-defined skill structures, LLM agents generate these contracts and corresponding motion templates for different objects and task contexts. These contracts guide the perception module to ground task-relevant 3D primitives from observations, which are then used to instantiate reusable motion templates stored in a skill library. We evaluate ManiSkillFormer on Galaxea R1-Lite dual-arm robot across three settings: zero-shot pick-and-place over 8 object categories with 30 different instances, functional manipulation tasks including unscrewing, pouring, pressing, and folding, and 3 long-horizon tasks. ManiSkillFormer achieves higher average success rates than the evaluated baselines and two ablated pipelines: 88.24% for demonstration-free pick-and-place, 75.00% average success on functional manipulation and 50--80% completion rates across the long-horizon tasks. These results show that our design enables composable and reusable manipulation across objects and tasks without per-object policy fine-tuning or additional robot demonstrations.
Robotics (cs.RO)
arXiv:2609.16331 [cs.RO]
(or arXiv:2609.16331v1 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2609.16331 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 279 words
Related coverage
VLBiMan++: Expanding the Generalization Boundary of Vision-Language Anchored One-Shot Bimanual Manipulation
arXiv · Sep 15, 2026
VOMMI: Collecting and Leveraging Portable Demonstrations for Mobile Manipulation
arXiv · Oct 7, 2026
[AINews] not much happened today
Latent Space · Sep 10, 2026
CORE: Common Outcome Regularities from Action-Free Visual Demonstrations for Robot Manipulation
arXiv · Aug 4, 2026