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

Younghyo Park, Antonia Bronars, Pulkit Agrawal

November 2026

TL;DR A controlled simulation study of how a robot's inertia relative to the object affects learning. RL copes with every mass ratio tested, but behavior cloning gets worse as the robot gets heavier, and a heavy robot pushing a light object needs 16x the demonstrations to match a light one.

What they do

In public simulation datasets, robot arms present a mean effective mass of 6 to 16 kg at contact, while the objects weigh 8 to 250 g. The paper asks whether this inertia mismatch costs robot learning anything. Controlled simulations vary robot and object mass with an identical free-space controller response and compare reinforcement learning (RL) with behavior cloning (BC). The paper also tests two ways to lower an existing arm’s effective mass at contact.

How it works

Point-mass study. A circular robot pushes a T. Both masses range from 0.1 to 5 kg, and PD gains scale with robot mass so free-space tracking is identical. Repeated over 9 controller settings.
RL expert, BC students. One PPO expert, conditioned on the masses and controller settings, generates 200 demonstrations per mass pair. A separate Transformer BC policy is trained for each pair.
Articulated robots. The same sweep runs on a simulated Flexiv arm and a 20-joint hand. Only the motors’ reflected inertia is scaled, so the robot’s weight stays fixed.
Two levers on a fixed arm. Either raise the workspace to postures with lower effective mass, or add a reward term that steers the RL demonstrator (a Franka reorienting a cube with a tool) toward low-inertia configurations.

Results

RL succeeds across the whole mass grid. BC success tracks the robot-to-object mass ratio and drops as the robot gets heavier.
A heavy robot pushing a light T needs 16 times the demonstrations to match a light robot pushing a heavy T with 100, and still trails at the largest dataset. Training error shows no mass trend, so poorer fitting does not explain the gap.
On the arm and the hand, BC success falls as reflected inertia grows, while RL still solves the task.
Raising the workspace to 0.40 m improved BC in all 36 evaluated settings. Low-inertia demonstrations improved most of 16 settings even though their training error was higher in all 16.

Limitations

All experiments are in simulation. Both levers also change the demonstrator’s strategy, so they show that a lever works without isolating inertia as the cause. Retargeting human demonstrations, teleoperation interfaces and sim-to-real effects are left open.

Videos

Figures

From Figure 1: Robot arms can far outweigh the everyday objects they manipulate. Can robot learning make up for this ... — From Figure 1: Robot arms can far outweigh the everyday objects they manipulate. Can robot learning make up for this mismatch with enough data, or does it carry a hidden cost?

From Figure 6: BC arrival success (%) on T-pushing for each robot and object mass, with one policy trained on 200 dem... — From Figure 6: BC arrival success (%) on T-pushing for each robot and object mass, with one policy trained on 200 demonstrations per cell. Success is highest when a light robot pushes a heavy T.

From Figure 7: BC success against the number of demonstrations for a light robot pushing a heavy T and a heavy robot ... — From Figure 7: BC success against the number of demonstrations for a light robot pushing a heavy T and a heavy robot pushing a light T. More data narrows the gap but does not close it.

From Figure 10: Two workspace placements for the same arm, on the ground (red, z = 0) and elevated to 0.40 m (blue), ... — From Figure 10: Two workspace placements for the same arm, on the ground (red, z = 0) and elevated to 0.40 m (blue), where the arm presents less inertia along planar push directions.

Abstract

Conventional robot arms can be extremely heavy compared with the everyday objects they manipulate, creating a significant inertia mismatch at contact. Even a small accidental bump from a heavy arm can knock a light object out of place. Modern learning methods have nevertheless demonstrated a wide range of successful manipulation behaviors on these platforms, suggesting that even extreme inertia mismatch can be overcome with sufficient data. However, it remains unclear whether high inertia imposes a hidden cost on learning, potentially worsening scaling with data or lowering the performance ceiling of modern robot learning methods. We investigate this question through controlled simulation experiments that vary robot and object masses. We find that reinforcement learning succeeds across a broad range of mass combinations, while behavior cloning learns more reliable policies when a lighter robot manipulates a heavier object. Heavier robots require more demonstrations to attain the same performance as lighter robots. Finally, because new hardware is rarely an option, we explore ways to lower an existing arm's inertia at contact with existing hardware, either by choosing a lower-inertia workspace before collecting data or by steering the demonstrator toward lower-inertia configurations during the task. In both cases, the lower-inertia demonstrations are imitated more reliably. Project website: https://younghyopark.me/inertia-mismatch/

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong