Sumo: Dynamic and Generalizable Whole-Body Loco-Manipulation

TL;DR

Sumo method enables quadruped robots to dynamically manipulate through sample-based planning, solving multi-task problems.

cs.RO 🔴 Advanced 2026-04-10 23 views
John Z. Zhang Maks Sorokin Jan Brüdigam Brandon Hung Stephen Phillips Dmitry Yershov Farzad Niroui Tong Zhao Leonor Fermoselle Xinghao Zhu Chao Cao Duy Ta Tao Pang Jiuguang Wang Preston Culbertson Zachary Manchester Simon Le Cléac'h
robotics dynamic manipulation reinforcement learning sample-based planning quadruped robots

Key Findings

Methodology

The study introduces a hierarchical control framework called Sumo, combining a pre-trained whole-body control policy with real-time sample-based planning. By adjusting the cost function at test time, Sumo adapts to different tasks and objects without additional training. The method uses a reinforcement learning-trained whole-body control policy to stabilize dynamics and adjusts in real-time through sample planning.

Key Results

  • Experiments on the Spot robot show that Sumo can upright tires heavier than the robot and drag larger obstacles, significantly improving success rates.
  • In simulation, Sumo successfully applies to humanoid robot tasks like door opening and table pushing, demonstrating broad adaptability.
  • Sumo adapts to new tasks by changing the object model or cost function at test time, without retraining.

Significance

This research is significant in the field of dynamic manipulation, addressing the challenge of robots coordinating their entire bodies to manipulate large or complex objects. The Sumo method offers new possibilities for real-world applications, especially in scenarios requiring flexible manipulation and dynamic adaptation.

Technical Contribution

Sumo's technical contributions lie in its hierarchical structure, combining the strengths of reinforcement learning and sample planning to provide stable and flexible manipulation capabilities. By reducing the action space dimension for online planning, Sumo enhances sample efficiency and maintains stability in unstable dynamic environments.

Novelty

Sumo is the first framework to combine sample planning with a pre-trained whole-body control policy for dynamic manipulation. Unlike existing methods, Sumo does not require fixed contact modes and can flexibly adapt to various tasks.

Limitations

  • Sumo may perform poorly with extremely complex object geometries, requiring more precise models for accurate dynamic predictions.
  • The computational cost of sample planning is high in high-dimensional action spaces.
  • Further research is needed to implement it on a broader range of robotic platforms.

Future Work

Future research could explore Sumo's application on different robotic platforms, optimize the computational efficiency of sample planning, and extend its adaptability in complex dynamic environments.

AI Executive Summary

Dynamic manipulation in robotics has long been a challenge, especially when robots need to handle objects larger than themselves. Traditional methods often rely on static environments or require complex pre-trained models, making them difficult to apply flexibly in the real world.

The Sumo method offers a new solution by combining a pre-trained whole-body control policy with real-time sample-based planning. This approach adapts to different tasks and objects without additional training. By adjusting the cost function at test time, Sumo demonstrates its adaptability across various dynamic manipulation tasks.

Experimental results show that Sumo successfully completes several complex tasks on the Spot quadruped robot, such as uprighting tires and dragging obstacles. In simulation, Sumo is also successfully applied to humanoid robot tasks. This research opens new possibilities for real-world robotic applications, although challenges remain in handling extremely complex objects.

Deep Analysis

Background

Research in dynamic manipulation has faced challenges, particularly in scenarios requiring full-body coordination. Traditional methods often rely on static environments or complex pre-trained models, making them difficult to adapt to changing real-world scenarios. Recent advances in reinforcement learning and sample planning have made significant progress in robotics, but stability in high-dimensional dynamic environments remains an issue.

Core Problem

Robots face issues of stability and flexibility when manipulating large or complex objects in dynamic environments. Existing methods often require complex pre-training or fixed contact modes, making them difficult to apply in diverse real-world scenarios.

Innovation

The Sumo method offers a new solution by combining pre-trained whole-body control strategies with real-time sample planning. Its innovations include not requiring fixed contact modes, flexibly adapting to various tasks, and improving sample efficiency by reducing the action space dimension for online planning.

Methodology

  • �� Use reinforcement learning-trained whole-body control policy to stabilize dynamics
  • �� Adjust in real-time through sample planning
  • �� Modify cost function at test time to adapt to different tasks
  • �� Combine low-dimensional action space with efficient sample planning

Experiments

Experiments were conducted on the Spot quadruped robot and G1 humanoid robot, testing Sumo's performance in various dynamic manipulation tasks. Benchmarks included uprighting tires and dragging obstacles, evaluating success rates and adaptability.

Results

Results show that Sumo performs excellently across multiple tasks, significantly improving success rates. Particularly in uprighting tires heavier than the robot and dragging larger obstacles, Sumo demonstrates strong adaptability.

Applications

Sumo can be widely applied in scenarios requiring flexible manipulation and dynamic adaptation, such as industrial robots and rescue robots. Its characteristic of not requiring additional training gives it significant advantages in changing environments.

Limitations & Outlook

While Sumo performs well across various tasks, it may perform poorly with extremely complex object geometries. Additionally, the computational cost of sample planning is high and needs further optimization.

Plain Language Accessible to non-experts

Imagine a robot trying to move a tire larger than itself. Traditional methods are like having the robot push the tire along a fixed track, but if the track is wrong, the tire won't move. The Sumo method is like giving the robot a flexible steering wheel, allowing it to adjust direction based on the tire's size and shape. This way, the robot can smoothly maneuver the tire in various situations without needing pre-set tracks.

ELI14 Explained like you're 14

Imagine you're playing a game where you control a robot to push a big tire. Traditional methods are like you can only push the tire along a fixed path, but if the path is wrong, the tire won't move. The Sumo method is like giving you a super flexible joystick, allowing you to adjust direction based on the tire's size and shape. This way, you can easily maneuver the tire in the game without needing pre-set paths. Isn't that cool?

Glossary

Reinforcement Learning

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

Used to train whole-body control strategies to stabilize dynamics.

Sample-Based Planning

A method that finds optimal paths by sampling and evaluating different strategies.

Used for real-time adjustment of robot manipulation strategies.

Whole-Body Control Policy

A strategy that controls the entire body's movements of a robot.

Used to stabilize the robot's performance in dynamic environments.

Dynamic Manipulation

The ability to manipulate objects while in motion.

The core problem studied in the research.

Quadruped Robot

A robot with four legs, typically used for complex terrains.

The robotic platform used in experiments.

Open Questions Unanswered questions from this research

  • 1 How to maintain Sumo's stability with extremely complex object geometries?
  • 2 How to further optimize the computational cost of sample planning for broader application scenarios?

Applications

Immediate Applications

Industrial Robots

The Sumo method can be used in industrial robots to help them flexibly manipulate large objects in dynamic environments, improving production efficiency.

Long-term Vision

Rescue Robots

The Sumo method can be applied to rescue robots, helping them manipulate obstacles in complex environments, enhancing rescue capabilities.

Abstract

This paper presents a sim-to-real approach that enables legged robots to dynamically manipulate large and heavy objects with whole-body dexterity. Our key insight is that by performing test-time steering of a pre-trained whole-body control policy with a sample-based planner, we can enable these robots to solve a variety of dynamic loco-manipulation tasks. Interestingly, we find our method generalizes to a diverse set of objects and tasks with no additional tuning or training, and can be further enhanced by flexibly adjusting the cost function at test time. We demonstrate the capabilities of our approach through a variety of challenging loco-manipulation tasks on a Spot quadruped robot in the real world, including uprighting a tire heavier than the robot's nominal lifting capacity and dragging a crowd-control barrier larger and taller than the robot itself. Additionally, we show that the same approach can be generalized to humanoid loco-manipulation tasks, such as opening a door and pushing a table, in simulation. Project code and videos are available at https://sumo.rai-inst.com/.

cs.RO