DAMAGESIM: Simulator-Agnostic Damage Modeling

• DAMAGESIM is a simulator-agnostic damage modeling layer that converts contact forces, temperature variations, and fluid interactions into cumulative, continuous damage signals.
• The framework augments standard POMDPs with an explicit health state, enabling real-time damage feedback and improved task-completion strategies in controlled environments.
• It supports multi-scale analysis with distinct DEMs for mechanical, thermal, and fluid damage, and finds applications in robotics, STEM disciplines, geomechanics, and structural health monitoring.

DAMAGESIM is the simulator-agnostic damage modeling layer inside OopsieVerse. It 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, and it augments a standard task-completion-centric POMDP with a health state that can be used in observations, rewards, costs, and termination conditions (Balaji et al., 30 Jun 2026). In a broader usage across recent literature, “DAMAGESIM” also appears as a capability label for damage simulation and damage-aware inference in domains including scanning transmission electron microscopy, robotic manipulation, geomechanics, hydraulic hazard reconstruction, and structural health monitoring (Rad et al., 30 Jun 2025, Fromm et al., 2016, Gaete et al., 2020, Raman et al., 2019, Marafini et al., 12 Mar 2026).

1. Formal damage-aware state representation

DAMAGESIM is defined over an environment containing objects

$\mathcal{E}=\{e_1,\ldots,e_N\},$

where each object may consist of multiple links $(l_1,\ldots,l_M)$. The simulator already provides each link’s physical state $s_{e_i}^{l_j}$, including pose, velocity, and contact-related information. DAMAGESIM augments this with a per-link health state $h_{e_i}^{l_j}$. Object-level health is defined conservatively as the minimum health across its links,

$$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$

so an object is considered effectively damaged once any critical link is badly damaged (Balaji et al., 30 Jun 2026).

At the task level, the starting point is a standard POMDP

$$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$

which is then augmented with a health state $h \in \mathcal{S}^h$, yielding a damage-aware POMDP with

$$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$

This augmentation does not alter the underlying simulator dynamics; it overlays a damage accounting process. Health is normalized onto a common scale

$h \in [0, h_{\max}], \qquad h_{\max}=100,$

where $100$ denotes fully undamaged and $(l_1,\ldots,l_M)$0 fully damaged. The full environment health state at time $(l_1,\ldots,l_M)$1 is

$(l_1,\ldots,l_M)$2

Health evolves through Damage Evaluation Models (DEMs). Each DEM produces an incremental damage quantity $(l_1,\ldots,l_M)$3, and object health is updated as

$(l_1,\ldots,l_M)$4

Damage is therefore continuous, accumulative, irreversible over an episode, and decomposable by modality. This representation distinguishes DAMAGESIM from binary failure labeling: it is an explicit state variable rather than a post hoc success/failure annotation.

2. Damage Evaluation Models

The core DEM families are mechanical, thermal, and fluid. Each maps simulator-native observables into thresholded, linearly accumulated damage increments. The framework is object-specific and link-specific through coefficients and thresholds (Balaji et al., 30 Jun 2026).

Damage category	Simulator signal	Increment law
Mechanical	Contact forces and link acceleration	$(l_1,\ldots,l_M)$5, then $(l_1,\ldots,l_M)$6
Thermal	Object temperature $(l_1,\ldots,l_M)$7	Piecewise thresholded hot/cold damage outside $(l_1,\ldots,l_M)$8
Fluid	Fluid-particle contact count $(l_1,\ldots,l_M)$9	$s_{e_i}^{l_j}$0

For mechanical damage, DAMAGESIM retrieves the set of contact forces $s_{e_i}^{l_j}$1 and the link acceleration $s_{e_i}^{l_j}$2. Each contact force is decomposed into components parallel and perpendicular to the acceleration direction: $s_{e_i}^{l_j}$3

$s_{e_i}^{l_j}$4

These are aggregated as

$s_{e_i}^{l_j}$5

The effective mechanical load is then

$s_{e_i}^{l_j}$6

and damage accumulates only beyond threshold: $s_{e_i}^{l_j}$7 The interpretation is explicit: $s_{e_i}^{l_j}$8 acts as a proxy for impulsive loading, while $s_{e_i}^{l_j}$9 approximates sustained, constrained, or shear/compressive loading.

For thermal damage, the model uses upper and lower safe temperature thresholds: $h_{e_i}^{l_j}$0 The appendix notes a code-level simplification using a single symmetric thermal scaling parameter $h_{e_i}^{l_j}$1.

For fluid damage, the raw signal is the amount of fluid particles in contact with each link, $h_{e_i}^{l_j}$2, and the model is

$h_{e_i}^{l_j}$3

The parameterization is deliberately object-specific. The paper gives examples such as a wineglass with

$h_{e_i}^{l_j}$4

a robot with

$h_{e_i}^{l_j}$5

and a box of crackers with

$h_{e_i}^{l_j}$6

These outputs are accumulated directly against the health scale $h_{e_i}^{l_j}$7–$h_{e_i}^{l_j}$8, not against calibrated SI damage units.

3. System architecture, simulator interface, and control formulations

Architecturally, DAMAGESIM is a plugin layer between the physics simulator and the task or learning stack. The underlying simulator produces $h_{e_i}^{l_j}$9, including kinematics, contacts, temperatures, and fluid interactions when available. DEMs read these signals at each timestep, compute modality-specific damages $$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$0, and update $$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$1, yielding the damage-aware state $$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$2. This health state can be exposed as observations, used to compute rewards or costs, visualized for teleoperation, or trigger termination when an object’s health reaches zero (Balaji et al., 30 Jun 2026).

The simulator-agnostic claim is narrow and concrete. DAMAGESIM requires access to common observables: contact forces per contact point or per link, link kinematics such as acceleration, object or link temperatures if thermal damage is needed, fluid-contact information if fluid damage is needed, object/link identity and articulation structure, and per-step state update hooks. On that basis, the same DEM logic can run across different simulators. In OmniGibson / Nvidia Omniverse, all three damage types are implemented. In RoboCasa / MuJoCo, only mechanical damage is instantiated, because thermal and fluid signals are not available in the same way.

The reporting structure is multi-scale. Damage is computed per link and per timestep, then summarized as per-modality damage, per-link health, per-object health, and full environment health. Thermal damage is object-level per timestep because OmniGibson exposes object temperature rather than link temperatures. This design supports both scalar control signals and object-centric attribution.

DAMAGESIM also provides formal control objectives. For reinforcement learning, the reward is

$$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$3

with $$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$4 controlling the strength of damage avoidance. In the appendix, the framework is also cast as a constrained MDP by defining

$$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$5

and solving

$$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$6

This makes DAMAGESIM usable not only as reward shaping, but also as an explicit safety cost.

4. OopsieVerse benchmark, evaluation protocol, and empirical uses

OopsieBench includes 32 household manipulation tasks across OmniGibson and RoboCasa, chosen to expose mechanical, thermal, and fluid damage modes. The benchmark distinguishes Task Completion Rate from Safe Task Completion Rate. Task completion ignores damage; safe completion requires both success and that all object health remain above $$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$7. The benchmark also reports Average Environment Health (Balaji et al., 30 Jun 2026).

The empirical uses are organized around four roles. First, DAMAGESIM provides real-time feedback during teleoperation through health bars initialized at 100 and optional red tinting on damaged objects. The paper reports safer demonstration collection on Shelve Cereal Box, Pick Egg, Add Firewood, Pour Water, and Wipe Counter when live health feedback is available.

Second, it supports damage-conditioned imitation learning and data curation. The reported failure mode is explicit: on Wipe Countertop, the unfiltered policy achieves perfect task completion but only $$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$8 safe completion. Policies trained with DAMAGESIM-informed data substantially improve safe completion with minimal drop in task completion.

Third, it acts as a reward signal for RL. On Shelve Cereal Box, DSRL improves safe task completion from $$h_{e_i} = \min\left(h_{e_i}^{l_1}, \ldots, h_{e_i}^{l_M}\right),$$9 to $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$0. In Move Cup of Water, PPO fine-tuning improves safe success from $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$1 to $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$2. In Place Plate, PPO trained from scratch with task reward alone yields almost $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$3 safe task completion, while adding the damage penalty yields over $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$4 safe task completion.

Fourth, it serves as an evaluation instrument for vision-language-action policies. Representative results in the table include: open microwave door in OmniGibson with $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$5 completion, $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$6 safe completion, and average health $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$7; ignite wood in OmniGibson with $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$8 completion, $$\mathcal{M}=(\mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \Omega, \mathcal{O}, \gamma),$$9 safe completion, and average health $h \in \mathcal{S}^h$0; Open Single Door in RoboCasa with $h \in \mathcal{S}^h$1 completion, $h \in \mathcal{S}^h$2 safe completion, and average health $h \in \mathcal{S}^h$3; and Turn On Stove in RoboCasa with $h \in \mathcal{S}^h$4 completion, $h \in \mathcal{S}^h$5 safe completion, and average health $h \in \mathcal{S}^h$6.

The sim-to-real result is similarly explicit. Aggregated across shelving a cereal box and pouring water on a Franka Panda, the damage-aware policy achieves $h \in \mathcal{S}^h$7 task completion, $h \in \mathcal{S}^h$8 safe completion, and $h \in \mathcal{S}^h$9 unsafe behavior rate, compared with $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$0 task completion, $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$1 safe completion, and $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$2 unsafe behavior rate for the damage-unaware policy. The paper reports this as a $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$3 reduction in unsafe behavior rate while preserving overall success.

5. Broader research interpretations of DAMAGESIM

Across the supplied literature, “DAMAGESIM” also denotes a broader damage-simulation capability rather than a single robotics framework. The recurring pattern is a domain-specific simulator or inference engine that converts physically meaningful fields or proxies into damage-relevant state variables, then uses those states for prediction, optimization, or classification (Rad et al., 30 Jun 2025, Fromm et al., 2016, Gaete et al., 2020, Raman et al., 2019, Marafini et al., 12 Mar 2026, Xu et al., 6 Aug 2025).

Domain	Paper	Role of damage simulation
STEM	“Fast Simulation of Damage Diffusion Distribution in Scanning Transmission Electron Microscopy” (Rad et al., 30 Jun 2025)	Event-based analytical superposition of per-dwell diffusion kernels for cumulative diffusion distribution
Robotic manipulation	“Physics-Based Damage-Aware Manipulation Strategy Planning Using Scene Dynamics Anticipation” (Fromm et al., 2016)	Physics simulation plus damage-oriented motion proxies such as weighted swept convex volume
Underground mining	“A fast algorithm of the shear-compression damage model for the simulation of block caving” (Gaete et al., 2020)	Quasi-static shear-compression gradient damage with relaxed alternating minimization
Dam-break hazard	“An investigation of the Brumadinho Dam Break with HEC RAS simulation” (Raman et al., 2019)	Screening-level inundation-footprint module rather than full end-to-end damage estimation
SHM benchmark	“Numerical benchmark for damage identification in Structural Health Monitoring” (Marafini et al., 12 Mar 2026)	Synthetic FAST/SLOW damage, EOVs, and sensor faults for algorithm validation
Sim-to-real SHM	“Bridging Simulation and Experiment: A Self-Supervised Domain Adaptation Framework for Concrete Damage Classification” (Xu et al., 6 Aug 2025)	Virtual testing platform plus domain adaptation for coda-wave damage classification

In STEM, the 2025 framework is explicitly “a specialized STEM diffusion solver rather than a universal beam-damage simulator.” It assumes a 2D diffusion model on an effectively infinite medium, uses analytical single-probe diffusion kernels $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$4, and forms the cumulative field

$$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$5

Its computational result is equally explicit: a full STEM scan with $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$6 probe positions and $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$7 frame time takes approximately 50 minutes in a Python implementation, whereas the C++ framework completes the same task in 16 seconds, a roughly $$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$8 speedup (Rad et al., 30 Jun 2025).

In robotic manipulation, the 2016 planner is “damage-aware” in a different sense. It does not predict fracture directly; it simulates scene dynamics and scores passive-object motion using proxies such as displacement, path length, maximum swept convex volume, and maximum weighted swept convex volume

$$s^{\textrm{DA}} \in \mathcal{S}^{\textrm{DA}}=\mathcal{S}\times\mathcal{S}^h.$$9

The sequence planner minimizes the sum of these per-step surrogate costs and, in the logistics setting, treats $h \in [0, h_{\max}], \qquad h_{\max}=100,$0 as a satisfactory threshold for “acceptable” versus “damaging” motion (Fromm et al., 2016).

In underground mining, the 2020 work uses a scalar damage field $h \in [0, h_{\max}], \qquad h_{\max}=100,$1 in a shear-compression gradient damage model, with a relaxed alternating update

$h \in [0, h_{\max}], \qquad h_{\max}=100,$2

applied when the staggered iterations oscillate. The reported effect is upward-propagating damage above the cave, which the authors emphasize as the physically relevant regime for caving and subsidence (Gaete et al., 2020).

In structural health monitoring, the 2026 numerical benchmark supplies 26280 hourly acquisitions over 2020–2022 for a fixed-fixed steel beam, combining environmental and operational variations, FAST damage, SLOW damage, and seven classes of sensor faults or malfunctions. That work functions as a practical DAMAGESIM-style dataset rather than a general constitutive simulator (Marafini et al., 12 Mar 2026). The 2025 concrete study extends the pattern to a virtual testing platform for coda-wave classification, where multiscale concrete degradation simulation is coupled to ultrasonic wave propagation, and unsupervised domain adaptation is used to bridge simulation and experiment (Xu et al., 6 Aug 2025).

6. Limitations, approximations, and recurring misconceptions

DAMAGESIM is explicitly an approximation rather than a full physical damage simulator. In OopsieVerse, mechanical fracture, burning, and fluid effects are reduced to thresholded linear proxies; fidelity is constrained by what the underlying physics engine exposes; object or link thresholds such as $h \in [0, h_{\max}], \qquad h_{\max}=100,$3, temperature thresholds, and fluid thresholds must be manually parameterized; and cross-simulator parity is incomplete because thermal and fluid damage are missing in RoboCasa / MuJoCo (Balaji et al., 30 Jun 2026).

The broader literature shows parallel limitations. The STEM framework is not a first-principles radiation-damage theory for all STEM damage channels; it operationalizes a previously derived diffusion-distribution model for a specific class of electron-beam damage (Rad et al., 30 Jun 2025). The manipulation planner does not compute impact impulse, contact force threshold, fracture, compliance, or actual object damage; it approximates damage severity from passive-object motion, especially weighted swept occupied volume (Fromm et al., 2016). The Brumadinho HEC-RAS study is best viewed as a hazard-module prototype: it models mudflow extent with standard 2D shallow-water equations and Manning friction, but omits tailings rheology, fragility, life-loss, and economic-loss components (Raman et al., 2019). The SHM benchmark is intentionally simple: dynamic response is SDOF, static response is a single Euler-Bernoulli midspan deflection, and the dataset is aimed at damage detection and robustness rather than spatial localization in complex MDOF systems (Marafini et al., 12 Mar 2026).

A recurrent misconception is to equate task completion, hazard routing, or anomaly detection with damage modeling. The supplied literature repeatedly separates those notions. In OopsieVerse, task success is insufficient if the robot damages itself or its surroundings (Balaji et al., 30 Jun 2026). In robotic manipulation, “damage-aware” refers to minimizing motion-derived risk proxies rather than simulating material failure (Fromm et al., 2016). In hydraulic reconstruction, inundation mapping is a foundation for later damage estimation, not the full pipeline (Raman et al., 2019). This suggests that DAMAGESIM is best understood not as a single constitutive model, but as a family of architectures in which damage-relevant state variables are made explicit, accumulated, and actionable within simulation.