arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

Heavy Robot Arms Can Master Lightweight Objects, but Learning Strategy Matters. | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
RoboticsEnglishPublished

Heavy Robot Arms Can Master Lightweight Objects, but Learning Strategy Matters.

November 4, 2026arXiv: 2611.03501v1

This huge inertial mismatch means that a small bump from a heavy arm can dislodge a light item, a risk that human hands avoid.

A diagram showing the large effective mass of a robot arm relative to the small mass of an everyday object it manipulates, highlighting the risk of accidental displacement.
Fig. 1. Illustration of the effective mass disparity between robot arms and everyday objects, emphasising that a heavy arm can easily dislodge a light item.Park, Younghyo, Bronars, Antonia, Agrawal, Pulkit · arXiv:2611.03501v1 · Fig. 1: · CC BY 4.0

Researchers tested whether advanced machine‑learning methods could overcome this problem. In a physics‑based simulation they varied the mass of both the robot and the held object. A Proximal Policy Optimization (PPO) reinforcement‑learning (RL) policy was first trained on a T‑shaped toy using 200 demonstrations and then fine‑tuned on a grid of robot and object masses.

The results were striking. RL succeeded in more than 85 % of the mass combinations tested, even when the robot was many times heavier than the object. By contrast, a transformer‑based behavior‑cloning (BC) policy performed reliably only when the robot’s effective mass was smaller than or comparable to the object’s. Adding more demonstrations narrowed the performance gap but never closed it. Thus, BC is particularly sensitive to high reflected inertia, whereas RL can adapt to a wide range of inertial conditions.

A graph depicting a decreasing trend in behavior‑cloning success rates with increasing reflected inertia of the robot, underscoring inertia as a hidden cost to learning. (1/2)A graph depicting a decreasing trend in behavior‑cloning success rates with increasing reflected inertia of the robot, underscoring inertia as a hidden cost to learning. (2/2)
(a). Showing that behavior‑cloning success declines as robot reflected inertia increases, even while the robot’s weight remains constant.Park, Younghyo, Bronars, Antonia, Agrawal, Pulkit · arXiv:2611.03501v1 · (a) · CC BY 4.0

The study shows that learning a manipulation policy on a heavy robot does not cripple the robot’s ability to handle light objects, but it does reveal a hidden cost of inertia for data‑efficient learning methods such as BC. Accounting for task‑space inertia when designing robots or collecting demonstration data can therefore reduce this cost and improve the reliability of learned skills.

Source paper

When Robots Outweigh Objects: How Inertia Mismatch Shapes Robot Policy Learning

Robotics
View on arXiv