TeraSim-World Overview

• TeraSim-World is a family of integrated simulation environments that model complex physical dynamics and support end-to-end autonomy and system evaluation.
• It groups three distinct implementations—TERA for excavation robotics, Terra for native 3D world modeling, and ns-3 TeraSim for full-stack THz networking—each coupling low-level physics with higher-level decision processes.
• Empirical validations across terrain response, 3D reconstruction, and network performance highlight practical tradeoffs between realism, scalability, and computational efficiency.

“TeraSim-World” is best treated as an Editor’s term for a family of simulation-oriented and world-modeling systems that aim to represent, evolve, and interrogate complex environments in a form suitable for end-to-end autonomy or systems evaluation. In the available arXiv literature, the term does not denote a single canonical software package; rather, it is most closely approximated by three distinct but related lines of work: TERA, a Unity3D- and AGX-based simulator for autonomous terrain excavation robotics (Aluckal et al., 2024); Terra, a native 3D world model that generates explorable environments in sparse point-latent space and decodes them into 3D Gaussians (Huang et al., 16 Oct 2025); and ns-3 TeraSim, a full-stack sub-terahertz and terahertz network simulation environment extended with realistic above-100 GHz channel models (Gargari et al., 2021). Taken together, these systems define a technical motif centered on integrated world representation, real-time or progressive environment evolution, and tight coupling between low-level physical structure and higher-level decision or protocol stacks.

1. Terminological scope and conceptual identity

The strongest literal evidence for “TeraSim-World” comes indirectly. TERA is described as a “TeraSim-World-like” environment because it is meant to be an integrated, extensible world model for robotics autonomy (Aluckal et al., 2024). Terra is described as “very much a TeraSim-World-like model” in spirit because it represents environments natively in 3D, generates them progressively, and supports exploration beyond initially observed regions (Huang et al., 16 Oct 2025). The ns-3 paper is described as directly relevant to “TeraSim / TeraSim-World-style full-stack simulation” because it extends THz networking simulation from a simple channel abstraction toward end-to-end protocol evaluation (Gargari et al., 2021).

System	Domain	Defining emphasis
TERA	Excavation robot autonomy	Unity3D + AGX, deformable terrain, ROS, multi-excavator simulation
Terra	Native 3D world modeling	Point latents, 3D Gaussians, exact multi-view consistency, exploration
ns-3 TeraSim	Above-100 GHz networking	Full-stack evaluation, directional antennas, HB/FS channel models

A common misconception is to read the expression as the name of a single monolithic simulator. The cited works instead indicate a looser constellation of systems that share an architectural ambition: each tries to make the “world” itself a computational object that can be simulated, rendered, sensed, or traversed without reducing the problem to a narrow subsystem. This suggests that “TeraSim-World” is best understood as a category of integrated simulation environments rather than as a standardized platform name.

2. Robotic excavation realization: TERA

TERA, short for Terrain Excavation Robot Autonomy, is a simulator specifically built to close a gap in excavation research: prior systems often emphasized either operator training or high-fidelity soil/contact physics, but not the simultaneous support for perception, control integration, multi-robot scalability, and real-time execution required for end-to-end autonomous excavation (Aluckal et al., 2024). At the system level, TERA is built on Unity3D and AGX Dynamics. Unity3D provides the visual scene, rigid-body simulation, and software environment for deploying the excavator, terrain, and sensors, while AGX is used as the interaction engine for terrain deformation and soil-tool/track contact. AGX is described as computationally efficient because it simulates detailed physics primarily where physical interaction occurs rather than across the entire scene at uniform high fidelity.

The excavator model is based on a real machine, the Takeuchi TB-235 compact excavator, and includes four main components: a 4-DOF manipulator, the cab, the base, and the tracks. The manipulator sits on a cab that can slew about the base, and the base is coupled to the tracks, giving the vehicle non-holonomic differential-drive-like motion. Tracks are modeled as custom AGX assets with parameters including thickness, material, number of links, and inter-link tension. The simulator preserves geometry and kinematic structure through a CAD-to-URDF workflow in which the excavator is designed in CAD, joints are established, dimensions are measured accurately, and the STEP model is converted to URDF through an Onshape-to-Robot pipeline for direct Unity import.

Terrain is represented as a 3D grid of cells containing material properties such as mass, compaction, and soil type. Different materials including soil, gravel, and sand are represented through parameters such as Young’s modulus, friction angle, and cohesion. For digging, the bucket and plow are modeled as AGX shovel elements with top-edge, bottom-edge, and cutting-edge geometry. When the bucket contacts the ground, the angle between the cutting edge and the terrain is evaluated, and soil deformation depends on that angle and the terrain’s shear properties. Excavated soil is converted into dynamic particles that interact with the terrain mesh, other particles, and the excavator itself. The simulator therefore models coupled excavation geometry, soil state, and vehicle motion rather than only mesh displacement.

TERA also incorporates detailed sensing and control interfaces. Through the Unity Sensors plugin, it supports RGB cameras, RGB-D cameras, IMUs, and LiDARs. Sensor resolution, acceleration bias, noise, pose offsets, and rotations are user-definable through a YAML configuration file that can be changed at launch time. Sensors can be attached to the chassis, arm, boom, and bucket. ROS integration includes publication of Odometry, JointState, Transform, and bucket mass information, and a custom ROS wrapper exposes a DeltaCAN control message in the normalized range $[-1, +1]$, emulating the excavator’s lever-to-CAN hardware interface. Multi-robot support is handled by YAML-generated multiple excavators with separate identifiers and namespaces.

A notable modeling contribution is the empirical time-varying actuation model for excavator joints:

$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$

where $i \in \{\text{boom, arm, bucket}\}$, $\omega^i_{ss}$ is the experimentally measured steady-state angular velocity, and $\eta^i$, $\beta^i$, and $\phi^i$ control oscillation frequency, decay, and phase delay. The reported experiments use $\eta^i = 20$, $\beta^i = 6$, and $\phi^i = 0$. This is intended to capture transient actuation behavior more faithfully than instantaneous velocity assignment.

3. Native 3D world modeling realization: Terra

Terra is a native 3D world model whose central claim is that a world model should evolve and generate in 3D directly rather than infer 3D only as a side effect of 2D synthesis (Huang et al., 16 Oct 2025). Its internal representation is a sparse set of point latents, written as

$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$0

so that each latent point contains a 3D position and a feature vector. These point latents lie on or near scene surfaces and are more compact than the original colored point cloud.

The architecture consists of two main components. The first is the Point-to-Gaussian VAE (P2G-VAE), which encodes a colored point cloud $$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$1 into a latent point set and decodes that latent set into 3D Gaussian primitives for rasterization. The asymmetry is fundamental: latent points are sparse and modeling-friendly, while decoded Gaussians are rendering-friendly and jointly represent geometry and appearance. Because naive Gaussian regularization of coordinates would destroy geometry, Terra introduces robust position perturbation, adding Gaussian noise to point positions in a locality-preserving way so that the decoder becomes robust to the noise encountered during generation.

The decoder performs adaptive upsampling and refinement. Each latent point expands into $$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$2 child points using learned queries, relative displacement prediction, and feature residuals; a refinement module then adjusts positions with learned offsets. The full VAE objective combines image-quality, geometric, Gaussian regularization, color-matching, and KL terms:

$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$3

The second component is SPFlow, a flow-matching model defined directly over latent point space. It starts from Gaussian noise and learns the transport trajectory to clean latent points:

$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$4

with a training objective

$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$5

A defining modeling choice is that positions and features are diffused jointly, coupling geometry and texture. Because point sets are unordered, Terra adds distance-aware trajectory smoothing, solved with the Jonker–Volgenant algorithm, to reduce unnecessary complexity in transport paths between noise samples and latent points.

Terra’s “explorable” property is implemented as outpainting in point latent space. Training proceeds in stages—reconstruction, unconditional generative pretraining, and masked conditional generation—with conditioning patterns that include cropping, uniform sampling, and combined crop + sampling. Conditioned latent points are concatenated with noisy latents and kept fixed during diffusion. The result is progressive generation in point latent space: the model can repeatedly extend a known boundary, choose a direction, step forward, and generate the next partial scene conditioned on previously generated regions.

The principal representational consequence is exact multi-view consistency. Once the internal state is a 3D scene encoded by point latents and decoded as 3D Gaussians, rendering from an arbitrary camera pose is standard 3D rasterization rather than a new view-conditioned generation step.

4. Full-stack networking realization: ns-3 TeraSim

In the networking literature, TeraSim refers to an ns-3 extension for macro wireless networks in the sub-terahertz and terahertz spectrum, and the above-100 GHz indoor study extends it beyond a mostly LOS-only propagation assumption by implementing two realistic 140 GHz channel models: a Hybrid (HB) ray-tracing/statistical model and a Fully Stochastic (FS) spatial channel model (Gargari et al., 2021). The significance of this work is that it frames TeraSim not as a channel-only environment but as a full-stack evaluation platform in which PHY, MAC, transport, application traffic, directional antennas, mobility, and LOS/NLOS transitions interact.

The HB model combines a ray-traced component and a stochastic component,

$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$6

and is tied to wideband measurements at 130–143 GHz in a conference room. It uses Fresnel equations, geometrical data, and wall reflection/scattering path loss, with ray-traced multipath components generated offline using the Q-D Channel RT tool and loaded at runtime via the qd-channel module. The FS model is a 3GPP-like indoor spatial channel model derived from measurement campaigns at 28 GHz and 140 GHz in an office environment, organized through Time Clusters (TCs) and Spatial Lobes (SLs). The directional CIR is written as

$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$7

Within TeraSim, the THzDirectionalAntenna class uses a cosine antenna model and can be static or rotating, with configurable rotation speed, initial phase, maximum gain, and beamwidth. The packet-processing flow is explicitly layered: THzSpectrumValueFactory creates a waveform or packet object; THzChannel checks node orientation; THzDirectionalAntenna provides transmit and receive gain; THzSpectrumPropagationLoss computes received power from antenna gain and channel impulse response; and the packet with computed power is passed to the receiver PHY. The channel type is selected through the ChannelType attribute in THzChannel.

The reported indoor scenario matches the HB measurement room and uses 140 GHz carrier frequency, 32 GHz bandwidth, mobility from point A to point B, 25 dBi maximum gain at both transmitter and receiver, receiver HPBW from 2° to 10°, -160 dBm noise floor, UDP transport, constant bitrate traffic, and source rates of 4, 12, 24, and 60 Gbit/s. The paper’s principal point is methodological: above-100 GHz network evaluation cannot separate channel realism from antenna directivity and stack-level behavior.

5. Empirical behavior and validation

TERA validates realism along three axes: terrain response, actuation response, and navigation/path similarity (Aluckal et al., 2024). For deformable terrain, the reported experiments compare dirt, gravel, and sand, assigning Young’s moduli of 6.5 MPa, 4.6 MPa, and 4 MPa, respectively. Velocity and acceleration traces show reduced variance over dirt and gravel, while sand causes a sharper speed decrease because lower normal forces cause the excavator to sink. The paper also presents terrain elevation after three digging actions, showing deformation from both bucket excavation and track traversal. For actuation realism, boom and arm angular velocity on the real excavator and in simulation are compared under a 60% input, and the time-varying response tracks the real velocity more closely than a simplistic instantaneous response. For end-to-end control validation, joystick inputs recorded on the real machine are replayed in simulation via ROS bag; the real trajectory is measured using two GPS receivers and converted from geodetic coordinates to local ENU coordinates. The real excavator traveled 57.39 m, the simulated path was 50.47 m, and the reported RMSE is 1.376 m.

Terra evaluates reconstruction, unconditional generation, and image-conditioned generation on ScanNet v2, using 1,513 RGBD scans with 958 training and 243 validation scenes (Huang et al., 16 Oct 2025). In reconstruction, baselines include PixelSplat, MVSplat, Prometheus, and Can3Tok$$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$8; metrics include PSNR, SSIM, LPIPS, Abs. Rel., RMSE, and $$\omega^i(t) = \omega^i_{ss}\left(1+\sin(\eta^i t + \phi^{i})e^{-\beta^i t}\right),$$9. Terra reports PSNR 19.742, SSIM 0.753, Abs. Rel. 0.026, RMSE 0.137, and $i \in \{\text{boom, arm, bucket}\}$0, with the best overall numbers except LPIPS. In unconditional generation, compared with Prometheus and Trellis, Terra reports P-FID 8.79 and P-KID 1.745, substantially outperforming Prometheus at 32.35 and 12.481 and Trellis at 19.62 and 7.658 on point-cloud geometry metrics. Its image metrics are worse than Prometheus, at FID 307.2 and KID 18.919 versus 263.3 and 10.726, which the paper interprets as a tradeoff in favor of stronger 3D structural fidelity. In image-conditioned generation, Terra reports Chamfer distance 0.217 and EMD 0.474, again emphasizing geometry and multi-view consistency.

The ns-3 TeraSim channel-model comparison evaluates strongest-path received power, angle-of-arrival structure, throughput, latency, beamwidth sensitivity, and received power during mobility (Gargari et al., 2021). In LOS, HB tends to produce stronger multipath components than FS, with less than about a 5 dB offset; in NLOS, the two are closer, though HB slightly outperforms FS in the third quartile. HB concentrates more MPCs near the LOS direction, which matters for narrow-beam alignment. At low source rates, all channel models behave similarly; at high source rate, especially 60 Gbit/s, HB achieves higher throughput and lower latency, and in some cases is the only model that reaches throughput close to the offered load. Beamwidth strongly modulates these differences: with 10° beams, HB and FS become more similar; with 2° beams, differences grow significantly. During mobility, the link is NLOS before $i \in \{\text{boom, arm, bucket}\}$1 s and LOS afterward; received-power differences are larger in NLOS and shrink when the direct path dominates.

6. Constraints, tradeoffs, and synthesis

Each realization of the TeraSim-World motif has explicit limitations. In TERA, the low-level dynamics engine is not directly exposed, hydraulic fluid and some inter-link friction effects are omitted, AGX computation scales with terrain size and number of dynamic particles, and photo-realistic camera rendering is expensive (Aluckal et al., 2024). In Terra, evaluation is concentrated on indoor ScanNet scenes; scalability to far larger, more dynamic, outdoor, or long-horizon interactive worlds remains unproven, image-quality metrics are not always best, and the system still depends on point-cloud preprocessing and training stages rather than being a fully end-to-end open-world simulator (Huang et al., 16 Oct 2025). In ns-3 TeraSim, the comparison is restricted to an indoor office/meeting-room scenario, 140 GHz carrier, 32 GHz bandwidth, UDP constant-bitrate traffic, and one family of antenna settings; the literature also notes that current above-100 GHz models lack a good temporal characterization of LOS/NLOS transitions (Gargari et al., 2021).

These limitations clarify the scope of the concept. “TeraSim-World” does not denote universal simulation completeness. Instead, it suggests an architectural program in which a world representation is made sufficiently explicit, configurable, and operational that higher-level stacks can interact with it without collapsing physical structure into a minimal surrogate. In TERA, that means coupling excavator geometry, terrain deformation, sensor streams, and ROS control. In Terra, it means storing the scene in native 3D latent form so that exploration and rendering arise from the same internal state. In ns-3 TeraSim, it means connecting realistic spatial channels and directional antennas to throughput and latency outcomes at the protocol level. The common thread is therefore not application domain but the insistence that environment structure, interaction dynamics, and downstream autonomy or networking behavior must be modeled as a single technical system rather than as isolated modules.