CoDex: Learning Compositional Dexterous Functional Manipulation without Demonstrations

TL;DR

CoDex framework achieves complex dexterous functional manipulation without demonstrations, with a 73% success rate.

cs.RO 🔴 Advanced 2026-07-01 44 views
Bowen Jiang William Painter Reger Roberto Martin-Martin
robotics dexterous manipulation no-demonstration learning vision-language models reinforcement learning

Key Findings

Methodology

CoDex employs vision-language models to generate semantic constraints, combining analytic optimization and reinforcement learning to autonomously discover complex dexterous functional manipulation strategies. The method infers semantic constraints directly from tasks and scenes, generating functional grasp candidates refined into full policies through reinforcement learning.

Key Results

  • CoDex achieved a 73% average success rate across six tasks in real-world settings, significantly outperforming baseline methods.
  • VLM-CEM generated global constraints that were both semantically and physically valid, outperforming existing methods.
  • The reinforcement learning stage significantly improved functional manipulation success rates by over 40%.

Significance

CoDex addresses the challenge of integrating semantic understanding with physical dexterity in functional manipulation, providing a novel approach for autonomous robot learning without demonstrations.

Technical Contribution

CoDex uniquely combines vision-language models with analytic optimization and reinforcement learning, achieving complex functional manipulation strategy learning without demonstrations, offering new theoretical guarantees and engineering possibilities.

Novelty

CoDex is the first to achieve complex functional manipulation guided by semantic constraints without demonstrations, differing from data-intensive methods reliant on demonstrations or task-specific training.

Limitations

  • Failures on high-precision objects due to sim-to-real discrepancies.
  • Occasional spatial misalignment in global constraint generation.

Future Work

Future research could explore enhanced domain randomization and closed-loop policies to improve success rates on high-precision objects.

AI Executive Summary

CoDex is an innovative framework designed to tackle the complex challenges of dexterous functional manipulation in robotics, especially without human demonstrations. Existing methods often rely on extensive demonstration data or task-specific training, whereas CoDex uses vision-language models to generate semantic constraints, combining analytic optimization and reinforcement learning to autonomously discover complex functional manipulation strategies.

In experiments, CoDex demonstrated outstanding performance across six different tasks, achieving an average success rate of 73%. These tasks, including spraying plants and using a hot glue gun, involved complex coordination of internal mechanisms and external control. CoDex guides analytic optimization to generate functional grasp candidates, which are refined into full policies through reinforcement learning.

While CoDex excels in many areas, challenges remain with high-precision objects, particularly due to sim-to-real discrepancies leading to operational failures. Future research could explore enhanced domain randomization and closed-loop policies to improve success rates on high-precision objects.

Deep Analysis

Background

Dexterous functional manipulation in robotics is a longstanding challenge involving the activation of internal mechanisms and control of external poses. Traditional methods rely on demonstration learning or task-specific training, which are data-intensive and have limited applicability.

Core Problem

CD-FOM tasks require bridging semantic understanding and physical dexterity, where robots must interpret task contexts and execute precise physical interactions, posing challenges for existing methods.

Innovation

CoDex generates semantic constraints through vision-language models, combining analytic optimization and reinforcement learning to achieve complex functional manipulation strategy learning without demonstrations, offering new theoretical guarantees and engineering possibilities.

Methodology

  • �� Vision-language models generate semantic constraints
  • �� Analytic optimization generates functional grasp candidates
  • �� Reinforcement learning refines into full policies
  • �� Tasks tested in simulation and real environments

Experiments

Experiments conducted on a 7-DoF robot arm with a 16-DoF multi-fingered hand across six CD-FOM tasks validated CoDex's effectiveness in real-world settings.

Results

CoDex achieved a 73% average success rate across six tasks, significantly outperforming baseline methods. VLM-CEM generated global constraints that were both semantically and physically valid.

Applications

CoDex can be applied to complex robotic tasks such as dexterous operations in automated factories and daily tasks in service robots.

Limitations & Outlook

Failures on high-precision objects due to sim-to-real discrepancies. Occasional spatial misalignment in global constraint generation.

Plain Language Accessible to non-experts

Imagine a robot working in a kitchen, needing to grab a spray bottle and aim it at a plant. This process is like us using our hands to grab a bottle, aim at a target, and press the nozzle. CoDex acts like a smart assistant, learning to complete tasks without being taught step-by-step. It observes the environment and task, generating rules to guide its actions. This is like following a recipe in the kitchen, where we don't need someone to teach us every step.

ELI14 Explained like you're 14

Imagine you have a super cool robot helper that can do all sorts of things at home, like watering plants. This robot doesn't need you to tell it what to do; it learns how to use a spray bottle by itself. It's like a smart student who figures out how to grab the bottle, aim at the plant, and spray water. Isn't that awesome? This robot is as clever as you figuring out the best strategy in a game!

Glossary

Vision-Language Model

A model that combines visual and language information to understand semantic information in tasks and scenes.

Used to generate semantic constraints for tasks.

Analytic Constrained Optimization

A mathematical method for optimizing an objective function under specific constraints.

Used to generate functional grasp candidates.

Reinforcement Learning

A machine learning method that learns strategies through trial and error and reward mechanisms.

Used to refine grasp-move-actuate policies.

CD-FOM Task

Tasks requiring the activation of internal mechanisms and control of external poses.

Main application scenario for the CoDex framework.

Global Constraints

Spatial and functional requirements for the final pose of an object in a task.

Generated by VLM-CEM to guide policy learning.

Open Questions Unanswered questions from this research

  • 1 How to improve success rates on high-precision objects? Current methods face challenges with sim-to-real discrepancies.
  • 2 How to enhance the accuracy of semantic constraint generation, especially in complex scenarios?

Applications

Immediate Applications

Automated Factories

In industrial settings, robots can perform complex dexterous operations to improve production efficiency.

Long-term Vision

Service Robots

In homes and public spaces, robots can autonomously complete various tasks, enhancing quality of life.

Abstract

In this work, we study Compositional Dexterous Functional Object Manipulation (CD-FOM): tasks such as aiming and actuating a spray bottle on a plant or a glue gun on wood, which require both actuating an object's internal mechanism and controlling its pose to apply the object's function to the environment. These tasks pose significant challenges for robots due to the demanding integration of semantic understanding of the object's function, actuation mode, and application area with intricate physical dexterity to manage grasp stability, movement trajectory, and actuation. We introduce CoDex, a zero-demonstration framework that autonomously discovers CD-FOM manipulation strategies. CoDex uses vision-language models (VLMs) to infer semantic constraints from the task and scene. These constraints guide analytic constrained optimization to generate a short list of functional grasp candidates that can be efficiently refined with reinforcement learning to generate full grasp-move-actuate policies transferable from simulation to the real world. We evaluate CoDex on a 7-DoF robot arm with a 16-DoF multi-fingered hand across six CD-FOM tasks involving previously unseen objects with internal mechanisms, including spray bottles, hot glue guns, air dusters, flashlights, and pepper grinders, and their application to unseen target objects, showcasing its ability to autonomously discover and execute complex, physically viable dexterous behaviors without human demonstrations. More information at https://robin-lab.cs.utexas.edu/CoDex/.

cs.RO