OopsieVerse: A Safety Benchmark with Damage-Aware Simulation for Robot Manipulation

Arnav Balaji, Arpit Bahety, Sriniket Ambatipudi, Daniel Lam, Junhong Xu, Roberto Martín-Martín · · 2026

Summary

Abstract Summary

While robotic manipulation capabilities have advanced rapidly, physical safety remains a major barrier to deploying household robots: task success is insufficient if the robot damages itself or its surroundings. Simulation offers a harm-free alternative to costly and dangerous real-world training and evaluation, yet existing simulators lack general mechanisms to detect, quantify, and represent damage. To address this gap, we introduce OOPSIEVERSE, a unified simulation framework and benchmark for damage-aware household manipulation. OOPSIEVERSE provides damage as an explicit, physically-grounded, and taskagnostic signal by converting sources such as contact forces, temperature changes, and liquid interactions into corresponding mechanical, thermal or fluid damage. OOPSIEVERSE comprises two core elements: (1) DAMAGESIM, a simulator-agnostic framework for detecting and quantifying damage during navigation and manipulation, and (2) a suite of household tasks designed to evaluate common damage modes and distinguish between task completion and safe execution. We demonstrate the generality of our framework by instantiating DAMAGESIM in two simulators with different physics backends, OmniGibson (Nvidia Omniverse) and RoboCasa (MuJoCo). We further showcase the utility of OOPSIEVERSE across multiple use cases, including (1) guiding safer demonstration collection via real-time damage feedback, (2) learning safer manipulation policies through damage-conditioned imitation learning and reinforcement learning, (3) benchmarking the safety of state-of-the-art Vision Language Action policies, and (4) improving real-world safety of sim-to-real transferred policies. Together, our results highlight the potential of OOPSIEVERSE as an open-source foundation for systematic, scalable research on safe robot manipulation. For code and more information, please refer to https://robin-lab.cs.utexas.edu/oopsieverse/

Key Points

Proposes OopsieVerse for robotics.
Evaluated on real or simulated robotic tasks.
Primary contribution in ar-vr.

DamageSim

In this section we introduce DamageSim , our simulation-based damage evaluation framework. After introducing our conceptual framing for a damage-aware POMDP formulation, we describe how we compute mechanical, thermal, and fluid damage signals from the simulator.

OopsieBench

We introduce a suite of 32 task instances across the two simulators (OmniGibson/OmniVerse for B1K tasks and MuJoCo for RoboCasa tasks), corresponding to 21 unique task designs (see Fig. 4 ).

The suite is built to (i) expose policies to physically damaging failure modes that arise naturally in household manipulation, and (ii) make the safety distinction measurable : for most tasks there exists an easy, tempting strategy that completes the goal but risks damage, alongside a safer strategy that typically requires more careful interaction (e.g., ge...

OopsieBench — OopsieVerse: A Safety Benchmark with Damage-Aware Simulation for Robot Manipulation — Fig. 7 : Trajectory differences based on damage feedback . Trajectories generated by the policy trained on w/o health feedback dataset that is not aware of damage ( red ) constantly pushes towards fragile objects. In contrast, the policy trained with health-filtered episodes ( green ) more frequently directs its trajectories toward the paper bag—even though interaction is required—resulting in safer task completions.

Experiments

Our experiments are designed to demonstrate the utility and versatility of OopsieVerse as a safety framework for robotic manipulation. We validate our framework by demonstrating the following use cases across two simulators (OmniGibson and Robocasa): Does OopsieVerse provide a sufficiently accurate signal to collect safer data and/or to train safer policies with imitation learning (Sec. V-A )?

Can OopsieVerse ’s extended state be used to define safety-related reward to train/finetune policies with reinforcement learning (Sec.

Experiments — OopsieVerse: A Safety Benchmark with Damage-Aware Simulation for Robot Manipulation — Fig. 9 : Sim-to-real transfer of policies trained in OopsieVerse . We evaluate baseline IL policy (without health feedback) with the IL policy with health-filtered episodes on Panda robot. The baseline IL policy often performs unsafe behaviors like spilling water on the laptop or pushing a fragile bottle over the shelf. Whereas the damage-aware filtered-episode IL policy learns safer behaviors. The damage-aware IL policy results in a 60 % safer execution as compared to the baseline IL policy. This shows that OopsieVerse can enable transferring safer policies to the real world.

Sources and Demonstrations

Figures are reproduced from Balaji et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.