Discoverse: Cross-Domain Discovery Systems

• Discoverse is a recurring label for systems that unify heterogeneous discovery processes, from federated metadata harvesting to semantic interpretation across multiple domains.
• It employs varied methodologies such as centralized metadata indexing, multimodal data analysis, and simulation-based robotics to extract hidden structures in complex information environments.
• Applications include data discovery (Mercury), interactive human behavior analysis (DISCOVER), robotic simulation and benchmarking (DISCOVERSE/Discoverse-L), pharmaceutical co-scientist systems (DiscoVerse), and semantics-driven image compression (DISCOVER).

Discoverse is a name and naming pattern used in several distinct computational systems concerned with discovery, exploration, or synthesis. In the literature, it denotes a metadata-centric “discoverse” layer for federated repositories, an interactive framework for human behaviour analysis, a 3DGS-based robot simulation framework and its long-horizon benchmark, a multi-agent pharmaceutical co-scientist, and a semantics-driven versatile codec for compression and machine vision (Palanisamy et al., 2010, Schiller et al., 2024, Jia et al., 29 Jul 2025, Liu et al., 20 Nov 2025, Zheng et al., 23 Nov 2025, Liu et al., 2024). This suggests that the term is best understood not as a single software lineage, but as a recurring label for systems that unify heterogeneous repositories, modalities, or tasks.

1. Terminological scope

A common source of confusion is the assumption that “Discoverse” denotes one platform. The published record instead contains several unrelated systems that share a discovery-oriented naming logic. Some are exact matches in spelling, while others use closely related stylizations such as DISCOVER or DiscoVerse.

Designation	Domain	Core description
Mercury “discoverse” layer	Data discovery	virtual internet repository
DISCOVER	Human behaviour analysis	data-driven, interactive system
DISCOVERSE	Robot learning	3DGS-based simulation framework
Discoverse-L	Robotic benchmarking	long-horizon manipulation benchmark
DiscoVerse	Pharmaceutical R&D	multi-agent co-scientist
DISCOVER	Image compression	versatile codec

The shared semantic core is “discovery,” but the technical substrates differ sharply. In one lineage, discovery means federated metadata harvesting over distributed repositories; in another, it means multimodal scene exploration for social science; in another, long-horizon robotic manipulation under stage-aware evaluation; and in another, traceable synthesis over historical pharmaceutical archives. This suggests that “Discoverse” functions as a cross-domain motif for systems that expose hidden structure in complex information environments rather than as a single standardized architecture.

2. Discoverse as federated metadata discovery

In Mercury, “discoverse” refers to a virtual internet repository that provides unified discovery over many physically distributed data repositories while leaving data ownership and storage with the original providers (Palanisamy et al., 2010). Mercury is organized into three major functional components: a Harvester, an Indexing tool / Search server, and User interface and services. The architecture implements a federated discovery model with a centralized metadata index: distributed data providers remain autonomous; they expose metadata via files in public web or FTP directories or via custom export programs that extract metadata from existing databases into XML; Mercury periodically harvests these metadata artifacts, transforms them if needed, indexes them in Lucene/Solr, and exposes search interfaces and machine-readable services.

The system supports heterogeneous metadata, including XML, Z39.50, FGDC, Dublin Core, Darwin Core, EML, and ISO 19115 (Palanisamy et al., 2010). Search modes include simple full-text search, advanced fielded search with temporal and spatial constraints, and web browse tree search. Spatial queries are integrated with Google Maps, and the new version also supports RSS delivery of search results. Result presentation includes faceted filters for data providers, parameter, sensor, topic, and project, as well as sorting by index rank, period of record, source, and project.

The central design principle is “centralized metadata, distributed data.” Mercury builds a centralized repository of metadata only; data files remain with the original providers, and metadata records point back to those sources. The paper states that the new version provides orders of magnitude improvements in search speed compared to the previous proprietary implementation (Palanisamy et al., 2010). A plausible implication is that Mercury established an early architectural template for later “discoverse”-style infrastructures: harvest, normalize, index centrally, and preserve provider autonomy.

3. DISCOVER as an interactive system for human behaviour analysis

DISCOVER is presented as a data-driven, interactive software system for comprehensive observation, visualization, and exploration of human behaviour, with four exemplary workflows: Interactive Semantic Content Exploration, Visual Inspection, Aided Annotation, and Multimodal Scene Search (Schiller et al., 2024). Its stated objective is to streamline computational-driven data exploration for human behavior analysis and to democratize access to advanced computational methodologies for researchers who lack extensive technical proficiency.

The architecture is described as a client-server web system with three major layers: a front-end, a back-end / API layer, and a data and model processing layer (Schiller et al., 2024). The front-end provides a media viewer with integrated transcripts, interactive visualizations, annotation tools, and text-based interfaces for semantic queries or interactions with an AI assistant. The processing pipeline begins from raw multimodal data, such as a Zoom video of a parent–teacher conference, and includes audio extraction, automatic speech recognition, transcript segmentation, model inference over segments, and storage of segments, metadata, and model outputs.

The analytical layer supports speech and text models, nonverbal and social-signal models, and multimodal embeddings (Schiller et al., 2024). A typical embedding formulation is given as

$\mathbf{z}_i = f(x_i) \in \mathbb{R}^d,$

with similarity often measured by cosine similarity,

$$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$

For visualization, DISCOVER computes two-dimensional coordinates

$\mathbf{y}_i = g(\mathbf{z}_i) \in \mathbb{R}^2$

using techniques such as t-SNE or UMAP. This supports clustering similar scenes, visual exploration of behavioural regions, and interactive retrieval of semantically or multimodally similar segments.

The system’s workflows illustrate a tightly coupled loop between computation and interpretation. Interactive Semantic Content Exploration uses transcript-grounded assistant queries such as summarization, criteria generation, and targeted analysis. Visual Inspection aligns timelines, speaker segments, and optional nonverbal features. Aided Annotation presents model-suggested candidates for human confirmation or correction. Multimodal Scene Search retrieves scenes using embedding-based similarity and interactive plots (Schiller et al., 2024). This suggests that DISCOVER’s contribution lies less in a single model than in integrating ASR, embeddings, visualization, annotation, and assistant-mediated querying into one exploratory environment.

4. DISCOVERSE in robotics: simulation infrastructure and long-horizon benchmarking

In robotics, DISCOVERSE is defined as the first unified, modular, open-source 3DGS-based simulation framework for Real2Sim2Real robot learning (Jia et al., 29 Jul 2025). Its core architecture combines a 3D Gaussian Splatting renderer, MuJoCo physics, and a ROS2 interface. Interactive entities use a dual representation: 3DGS for visuals and mesh assets for collision and dynamics. The Real2Sim pipeline distinguishes a background node for non-interactive environment structure from interactive scene nodes for objects and robots, and it supports real-world captures, 3D AIGC assets, public 3D libraries, and multiple robot models.

The framework emphasizes hyper-realistic visual reconstruction and accurate physics. Background scenes are reconstructed from multi-view image capture, laser scanning, and COLMAP camera poses; interactive objects can come from scanners or from CLAY for non-Lambertian, thin, or hard-to-scan objects; relighting is handled by estimating an HDR environment map with DiffusionLight and using Blender for physically based rendering before Mesh–Gaussian transfer (Jia et al., 29 Jul 2025). Sensor support includes RGB, depth, LiDAR, tactile sensing via Tacchi, IMU, and joint and body states. The paper reports more than 100 FPS for LiDAR via a BVH-accelerated, native Gaussian ray-tracing framework, and 650 FPS for 5 cameras rendering RGB-D at 640×480 on a desktop configuration, about 3× faster than Isaac Lab under similar conditions (Jia et al., 29 Jul 2025).

The empirical focus is zero-shot Sim2Real transfer for imitation learning. On three contact-rich manipulation tasks—Close-Laptop, Push-Mouse, and Pick-Up-Kiwifruit—DISCOVERSE yields the best zero-shot Sim2Real results among the tested simulators for both ACT and Diffusion Policy (Jia et al., 29 Jul 2025). With augmentation, ACT trained in DISCOVERSE reaches an average real-world success rate of 86.5%, and Diffusion Policy reaches 86.0%.

A related but distinct construct is Discoverse-L, which is defined as a long-horizon manipulation benchmark built on DISCOVERSE and AIRBOT-Play (Liu et al., 20 Nov 2025). It contains three multi-stage tasks: Block Bridge (74 stages), Stack (18 stages), and Jujube-Cup (19 stages). For each task, the authors collect demonstrations, run a video-driven stage discovery pipeline with Gemini 2.5 Pro, and build a stage dictionary with stage-wise text triplets $(T_k^+, T_k^-, T_k^{h-})$. This stage structure supports both reward shaping and explicit evaluation of stage hallucination.

EvoVLA, the model introduced around Discoverse-L, addresses long-horizon failure modes with Stage-Aligned Reward (SAR), Pose-Based Object Exploration (POE), and Long-Horizon Memory (Liu et al., 20 Nov 2025). The benchmark defines success rate, sample efficiency, and hallucination rate, where hallucination rate measures the fraction of high VLM stage scores that are false positives relative to simulator ground-truth stage completion. On Discoverse-L, EvoVLA reaches 69.2% average success, improves average task success by 10.2 percentage points over OpenVLA-OFT, achieves one-and-a-half times better sample efficiency, and reduces stage hallucination from 38.5% to 14.8% (Liu et al., 20 Nov 2025). In real-world deployment on physical robots, it reaches an average success rate of 54.6% across four manipulation tasks. A common misconception is to equate Discoverse-L with DISCOVERSE itself; the literature distinguishes the former as a benchmark suite built on the latter.

5. DiscoVerse as a pharmaceutical co-scientist

In pharmaceutical R&D, DiscoVerse is a multi-agent co-scientist designed to support reverse translation over large historical archives (Zheng et al., 23 Nov 2025). The reported corpus is a Roche subset comprising 180 molecules, 15,762 PDF files, 872,453,585 BPE tokens, and more than four decades of research. Documents are assigned to internal unique molecule identifiers, and documents that mention multiple molecules are assigned to all relevant identifiers, producing a molecule-centric retrieval universe.

The architecture is modular and role-specialized. A Classification and Decomposition Agent classifies and decomposes the user query; three domain branches—preclinical, clinical, and strategic—each include a Decomposition Agent, Search Agent, Review Agent, and Research Agent; a Supervisor Agent integrates outputs across branches; and a Taxonomy Agent maps results into structured schemas from a schema library co-designed with Roche scientists and project leads (Zheng et al., 23 Nov 2025). This aligns the system with scientist workflows rather than with a generic QA pipeline.

The retrieval stack combines VLM-based OCR with hybrid search. PDFs are parsed with olmOCR; text is segmented into 512-word chunks with 64-word overlap and section-aware boundaries; embeddings are built with intfloat/multilingual-e5-large-instruct and BGE-M3; BM25 lexical search is included; and results from dense retrieval, multi-vector search, and lexical search are merged and deduplicated (Zheng et al., 23 Nov 2025). ChromaDB serves as the vector store, and MongoDB is used for document storage and metadata management. Review uses both a learned reranker and LLM-based relevance judgment, with reranker score at least 0.7 and positive LLM relevance judgment required.

The evaluation emphasizes expert review rather than automatic text metrics. Quantitative assessment focuses on seven benchmark queries, including first-in-human dose, route of administration, highest dose, dose with severe adverse events, efficacious dose, regimen, and margin of safety; discontinuation rationale and multi-phase hematotoxicity receive qualitative assessment (Zheng et al., 23 Nov 2025). Across seven benchmark queries covering the 180 molecules, DiscoVerse achieves near-perfect recall $(\geq 0.99)$ with moderate precision $(0.71\text{–}0.91)$. The paper emphasizes that false positives are rarely pure hallucinations; they are more often context errors such as preclinical-versus-clinical confusion, planned-versus-actual confusion, or phase mislabeling. This suggests that DiscoVerse is designed as a high-recall, source-linked assistant for expert review rather than as an autonomous decision maker.

6. DISCOVER as a semantics-driven codec

In image compression, DISCOVER stands for semantics DISentanglement and COmposition VERsatile codec, a framework designed to simultaneously enhance human-eye perception and machine vision tasks (Liu et al., 2024). The stated motivation is that learned image compression methods are often specialized either for human visual perception or for machine vision tasks, requiring retraining for new applications. DISCOVER addresses this by performing task-aware semantic analysis at the encoder and diffusion-based semantic composition at the decoder.

The encoder-side workflow begins with task-level label generation by a multimodal large model, specifically GPT-4o, followed by image-level localization with Grounding DINO (Liu et al., 2024). Grounding DINO provides filtered image-relevant labels and bounding boxes $\boldsymbol{l}_i$, which are used to disentangle the main codec latent into task-related regions. The codec follows an ELIC-like structure with encoder $g_a$, hyperprior encoder $h_a$, decoder $g_s$, and a side-information extraction module $$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$0. The paper writes

$$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$1

so that task-related latent regions can be transmitted with the global hyperprior while background regions may be omitted.

At the decoder, DISCOVER reconstructs a diffusion VAE latent rather than pixels directly: $$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$2 This latent conditions Stable Diffusion 2.1 through a ControlNet-style module. The reverse diffusion step is expressed as

$$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$3

after which the diffusion VAE decoder produces the reconstructed image $$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$4 (Liu et al., 2024). Training is two-stage: Stage I optimizes the full codec and ControlNet-style module; Stage II freezes encoder and hyperprior, applies random masks to the main latent, and adapts the decoder and control module to partial latents.

The reported machine-vision gains are substantial: BD-rate is approximately $$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$5 for detection on COCO, $$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$6 for segmentation on COCO, and $$s(i, j) = \frac{\mathbf{z}_i \cdot \mathbf{z}_j}{\|\mathbf{z}_i\| \, \|\mathbf{z}_j\|}.$$7 for classification on ImageNet relative to VTM-12.1 (Liu et al., 2024). For human perception, the full-bitstream version outperforms PerCo and ILLM on FID, KID, and DISTS on CLIC and Kodak. Here “discover” does not refer to search over external repositories; it refers to discovering, disentangling, and recomposing the semantic structure of an image so that one codec can serve both human observers and machine tasks. This broader pattern is consistent with the term’s repeated use across the literature: discovery is framed as the controlled unification of heterogeneous signals into an operational substrate for retrieval, interpretation, or action.