Zero-Shot Exploration: Transfer and Applications

• Zero-shot exploration is a transfer regime that leverages pretrained semantic, visual, and world-model representations to handle unseen classes, tasks, and environments.
• It spans multiple domains—from recognition and reward-free reinforcement learning to embodied navigation and multimodal reasoning—by converting prior knowledge into exploratory policies.
• Key insights include improved performance metrics and techniques addressing challenges like domain shift, descriptor noise, and sparse semantic representations for robust transfer.

Zero-shot exploration denotes a family of methods that must identify, reason about, or act toward previously unseen classes, tasks, or environments without task-specific labeled examples or environment-specific fine-tuning. In classical zero-shot recognition, the unknown lies in label space and transfer is mediated by semantic side information shared by seen and unseen classes; in reward-free reinforcement learning, the unknown lies in downstream reward functions and exploration must learn a transferable control representation; in embodied navigation, the unknown lies in the environment, the target location, or both, and exploration must be driven by pretrained semantic priors, online memory, and geometry; in multimodal reasoning, the unknown lies in latent visual, linguistic, or speaker-identity structure that must be interrogated at inference time (Fu et al., 2017, Sancaktar et al., 2022, Bagot et al., 29 Jun 2026, Yang et al., 2023).

1. Conceptual scope and formal definitions

In the zero-shot recognition literature, the canonical setup partitions classes into seen and unseen label sets, with training restricted to seen classes and testing on disjoint unseen classes. Visual inputs are represented by features $\theta(x)$, class semantics by descriptors $\phi(y)$, and prediction is performed by transferring knowledge through the shared semantic space. The same survey formalizes inductive zero-shot learning as training on seen classes only, transductive zero-shot learning as additionally exploiting unlabeled unseen data, closed-set ZSL as testing only on unseen classes, and generalized ZSL as predicting over $S_{\text{seen}} \cup S_{\text{unseen}}$ (Fu et al., 2017).

The phrase later broadened. In online transfer with Behavioral Foundation Models, zero-shot exploration denotes using a reward-free pretrained policy family to explore and infer an unknown black-box reward during transfer, rather than receiving a pre-collected reward dataset offline (Bagot et al., 29 Jun 2026). In structured world-model manipulation, it denotes an intrinsically motivated, task-agnostic phase in which an agent plans for predicted novelty, learns object-centric dynamics online, and subsequently solves downstream tasks without additional policy training (Sancaktar et al., 2022). In embodied navigation, it typically means searching for a target object or executing language-guided navigation in unseen environments without environment-specific training, using pretrained perception and LLMs as the main source of prior knowledge (Dorbala et al., 2023, Zhou et al., 2023).

A plausible implication is that “zero-shot exploration” is best understood not as a single algorithmic paradigm but as a transfer regime. What unifies the literature is the absence of target-specific supervision at deployment time and the need to convert pretrained structure—semantic descriptors, uncertainty estimates, world models, spatial memory, or language priors—into an exploration policy.

2. Semantic transfer and exploration of unseen concept spaces

The foundational zero-shot recognition formulation learns a compatibility function between visual features and class semantics, most commonly

$F(x, y) = \theta(x)^\top W\, \phi(y),$

with prediction by $y^* = \arg\max_y F(x,y)$. This paradigm underlies ALE, SJE, DeViSE, ESZSL, and related methods. The same literature also includes DAP and IAP attribute-based probabilistic models, multi-view CCA, deep visual-semantic embedding, transductive label propagation, and generative feature synthesis for generalized ZSL. Attributes, word vectors, text corpora, and ontologies serve as the transfer substrate, while calibrated stacking addresses the systematic bias toward seen classes in generalized ZSL (Fu et al., 2017).

Later work made the “exploration” aspect more explicit by perturbing or expanding the semantic space itself. “Creativity Inspired Zero-Shot Learning” introduces hallucinated class descriptions obtained by interpolating seen-class text embeddings and imposes a creativity-inspired loss that encourages generated features to deviate from seen classes while remaining realistic. On the hard CUB split, GAZSL rises from Top-1 10.3 to 14.4 and SUC AUC from 8.7 to 11.9 under this scheme, and on AwA2 the same idea raises Top-1 from 58.9 to 67.8 and the harmonic mean from 15.4 to 24.6 (Elhoseiny et al., 2019). This suggests that zero-shot exploration can be implemented directly in semantic latent space rather than only in physical or visual space.

“COMAE” extends this idea to zero-shot hashing by defining comprehensive attribute exploration through point-wise, pair-wise, and class-wise consistency constraints. It regresses attributes from an attribute prototype network, uses contrastive learning over continuous-valued attributes, and jointly aligns image representation, visual attributes, and hash codes. At 64 bits, it reports mAP@5000 of 0.5133 on AWA2, 0.1994 on CUB, and 0.1415 on SUN, outperforming several prior hashing baselines (Li et al., 2024). Here, exploration is not embodied motion but systematic exploitation of attribute locality and attribute context to support unseen-class retrieval.

The survey literature also identifies the principal technical obstacles that make semantic exploration difficult: projection domain shift, hubness in high-dimensional spaces, sparsity or noise in semantic descriptors, and evaluation inconsistencies caused by nonstandard splits and pretrained features (Fu et al., 2017). Those issues recur, in modified form, in later robotic and multimodal work.

3. Multimodal extensions beyond object recognition

In large vision-LLMs, zero-shot exploration has been recast as expanding the model’s visual attention beyond the narrow regions named in the original query. “Question-Driven Visual Exploration” generates four short, input-conditioned pre-questions from the task instruction and query, then conditions LVLM reasoning on those questions. On ScienceQA, InstructBLIP improves from 49.3 to 55.0 with GPT-generated questions; on SNLI-VE, it rises from 34.5 to 50.1; and on fine-grained classification the same framework improves Flowers102 from 59.5 to 67.8 and Oxford-IIIT Pet from 69.8 to 80.5 (Yang et al., 2023). In this usage, exploration is attention allocation: the model is prompted to inspect peripheral or subtle evidence before answering.

In instruction tuning, the term appears in cross-lingual form. “Deep Exploration of Cross-Lingual Zero-Shot Generalization in Instruction Tuning” trains on one language’s meta-dataset and evaluates held-out task clusters in the other language, with cross-lingual templates designed to align both linguistic surface form and instruction format. The paper reports average gains of 20.7% on English-target evaluation and 13.6% on Korean-target evaluation relative to baseline mT5, and argues that relevant data acquisition across languages can matter more than linguistic congruence with the unseen tasks (Han et al., 2024). Exploration here concerns unseen task formats and unseen instruction languages rather than unseen physical environments.

Even in speech synthesis, a related regime appears. In zero-shot multi-speaker TTS, the exploration problem becomes the selection of a conditioning representation that best preserves speaker identity for unseen voices. A YourTTS-based comparison of H/ASP, x-vector, and ECAPA-TDNN embeddings on 24 out-of-domain Czech radio speakers found subjective speaker-similarity scores of 47.28 ± 26.19 for H/ASP, 42.62 ± 26.52 for ECAPA-TDNN, and 40.98 ± 26.24 for x-vector, with H/ASP also strongest on most objective cosine-distance evaluations (Kunešová et al., 25 Jun 2025). This broadens the term further: zero-shot exploration can denote exploration of representation spaces whose geometry determines transfer quality.

4. Reward-free reinforcement learning and intrinsic exploration

In model-based manipulation, “Curious Exploration via Structured World Models Yields Zero-Shot Object Manipulation” defines zero-shot exploration as planning for novelty inside a learned object-centric world model. The environment is a fully observable MDP whose state factors into object-centric components and an agent state, and the dynamics are modeled by an ensemble of message-passing GNNs. Novelty is quantified by epistemic uncertainty via the trace of the covariance of next-state predictions across ensemble members, and action sequences are optimized with iCEM in an MPC loop (Sancaktar et al., 2022). After intrinsic exploration only, the same model solves pick-and-place, stacking, flipping, and throwing without any additional model or policy training. Representative results include success 0.70 ± 0.04 on “push all objects to one location” after 100 iterations versus 0.08 ± 0.01 for an MLP baseline, and 0.97 ± 0.01 pick-and-place success on 4 objects in Construction (Sancaktar et al., 2022).

A different but related line studies zero-shot exploration at transfer time rather than pretraining time. In “Exploration and Online Transfer with Behavioral Foundation Models,” transfer is framed as a bandit-like exploration–exploitation problem in the Universal Successor Features framework. The transfer-time algorithm maintains

$$V_t = \lambda I + \sum_{k=1}^t \phi(S_k)\phi(S_k)^\top,$$

builds a confidence ellipsoid for the unknown reward parameter, and chooses conditioning vectors using a UCB-style criterion over successor features. The paper emphasizes minimizing eigenvalues of the uncertainty matrix $V_t^{-1}$, measuring uncertainty reduction through $\log \det(V_t^{-1})$, and using the BFM itself as the generator of exploration policies (Bagot et al., 29 Jun 2026). This makes zero-shot exploration an online experiment-design problem over policy space.

“Epistemically-guided forward-backward exploration” applies a closely related idea to Forward-Backward representations. The successor measure is factorized as a forward representation and a backward representation, and exploration is driven by ensemble-based posterior uncertainty over the forward component, projected to predictive variance in the induced $Q$-function. The method selects the latent-conditioned greedy policy with maximal predictive variance and reports substantially better sample efficiency than uninformed Forward-Backward exploration and random actions across five DeepMind Control Suite domains (Urpí et al., 7 Jul 2025).

“Maximum Entropy Behavior Exploration for Sim2Real Zero-Shot Reinforcement Learning” addresses the same problem in quadrupedal control. FB-MEBE maximizes the entropy of the achieved behavior distribution in task-relevant descriptor spaces such as planar velocity or base orientation, and adds a regularization critic penalizing joint acceleration, action rate, and feet slippage. The reported effect is broader coverage, improved zero-shot downstream performance, and direct hardware deployment without fine-tuning (Hu et al., 26 Mar 2026). Across these works, zero-shot exploration becomes a question of which uncertainty, density, or successor-feature geometry should be flattened during reward-free data collection.

5. Embodied navigation, spatial reasoning, and persistent memory

The most visibly embodied use of the term concerns navigation in unseen environments. Several systems guide exploration directly from pretrained semantic priors. ESC converts open-vocabulary GLIP detections and commonsense language-model scores over rooms and co-occurring objects into soft logic predicates, then selects frontiers by minimizing weighted hinge-loss violations in a PSL model. On MP3D it reports SPL 14.2 versus 3.7 for CoW and SR 28.7% versus 7.4%, and on RoboTHOR it reports SPL 22.2 versus 16.9 and SR 38.1% versus 26.7% (Zhou et al., 2023). LGX instead prompts GPT-3 with YOLO object lists or BLIP captions to choose the next exploratory subgoal while GLIP performs open-vocabulary target grounding; on RoboTHOR, the YOLO→LLM + GLIP variant reaches SR 35.0% and SPL 21.9%, improving over OWL-CoW at SR 27.5% and SPL 17.2 (Dorbala et al., 2023).

Other approaches minimize dependence on explicit maps or heavy symbolic reasoning. ClipRover uses only monocular RGB for decision-making, slices each image into six tiles, computes CLIP-based navigability, target, and familiarity scores, and switches among basic navigation, look-around, and target-lock modes. In real-world trials it reports success rate ≈95%, GPU-parallel throughput ≈5.01 FPS on a Jetson Orin, and performance comparable to path planners that require prior maps or target knowledge (Zhang et al., 12 Feb 2025). A separate sim-to-real line trains a compact PPO policy on 2D LiDAR racing and transfers it zero-shot to real AMR exploration without maps, domain randomization, or fine-tuning; the same policy is reported to lap short racetracks ≈30% faster than a modified wall-following PID and to use under 25% of available CPU and memory on a Jetson Orin AGX (Sivashangaran et al., 2024).

Recent embodied systems have emphasized spatial memory and hierarchical reasoning. SSR-ZSON augments the TARE hierarchical exploration framework with viewpoint scoring that balances geometric coverage and semantic density, plus an LLM-based global guidance mechanism over sub-regions; it reports SR/SPL of 59.5/0.345 on MP3D and 65.7/0.391 on HM3D (Meng et al., 29 Sep 2025). AERR-Nav organizes behavior into Exploration, Recovery, and Reminiscing states, with fast-thinking and slow-thinking modes inside exploration; on HM3D it reports SR=72.3% and SPL=35.6%, and on MP3D SR=47.9% and SPL=18.7%, with especially large gains on inter-floor tasks (Huang et al., 18 Mar 2026). Spatial-VLN combines Spatial Perception Enhancement with Explored Multi-expert Reasoning and activates query-and-explore when waypoint-level and region-level experts disagree, achieving SR 33 and SPL 27.44 on VLN-CE with Deepseek-v3 and strong real-world transfer via value-based waypoint sampling (Yue et al., 19 Jan 2026). ApexNav similarly switches between semantic-driven and geometry-driven exploration based on the distribution of semantic evidence and adds target-centric semantic fusion to preserve long-term memory of the target and visually similar objects (Zhang et al., 20 Apr 2025).

Persistent spatial memory has become a central design variable. GSMem uses 3D Gaussian Splatting as a persistent memory that supports post-hoc re-observability, meaning the agent can later render photorealistic novel views from previously unoccupied viewpoints. Its hybrid exploration strategy combines VLM semantic scores with a 3DGS-based geometric coverage objective, and it reports 55.4/43.8 on Active EQA and 67.2/46.9 on GOAT-Bench, outperforming prior graph- and snapshot-based memories (Lu et al., 19 Mar 2026). VLN-Zero likewise separates a rapid exploration phase from a deployment phase, building a symbolic scene graph through structured VLM-guided search and later exploiting a cache of task-location trajectories; on R2R Val-Unseen it reports SR 42.4 and SPL 26.3, and on RxR Val-Unseen SR 30.8 and SPL 19.0, while using 55% fewer VLM calls on average than prior zero-shot models (Bhatt et al., 23 Sep 2025).

6. Evaluation regimes, recurring bottlenecks, and future directions

The survey tradition already established that zero-shot exploration cannot be evaluated meaningfully without specifying the transfer regime. For recognition, the critical distinctions are inductive versus transductive training, closed-set versus generalized ZSL, per-class versus per-image accuracy, and split integrity; generalized ZSL should report both seen and unseen performance rather than only unseen accuracy (Fu et al., 2017). For navigation and embodied reasoning, the standard metrics remain SR and SPL, sometimes augmented by path length, VLM-call counts, or task-specific language-based scores such as LLM-Match and LLM-Match SPL (Lu et al., 19 Mar 2026, Bhatt et al., 23 Sep 2025).

Across subfields, the principal failure modes are structurally similar even when the observation spaces differ. In semantic ZSL they appear as projection domain shift, hubness, descriptor noise, and seen-class bias (Fu et al., 2017). In world-model control they appear as compounding model errors, finite-horizon planning limits, and unreliable contact dynamics (Sancaktar et al., 2022). In embodied exploration they appear as prompt sensitivity, detector hallucination, sparse semantics, partial observability, map errors, and difficulty with stairs, doors, specularities, or dynamic clutter (Dorbala et al., 2023, Yue et al., 19 Jan 2026, Lu et al., 19 Mar 2026). In multimodal prompting they appear as over-exploration, noise from too many auxiliary questions, or bias introduced by language priors rather than visual evidence (Yang et al., 2023).

Several future directions recur. The zero-shot recognition survey emphasizes richer semantic spaces, compositional semantics, open-world operation, and stronger evaluation standards (Fu et al., 2017). Embodied systems increasingly point toward persistent memory, active view planning, uncertainty-aware scene representations, and integration of LLMs with explicit spatial structure rather than replacing geometry with language (Lu et al., 19 Mar 2026, Bhatt et al., 23 Sep 2025). Reward-free RL points toward tighter connections between exploration and transfer representations, especially when rewards are black-box and transfer is online rather than offline (Bagot et al., 29 Jun 2026, Urpí et al., 7 Jul 2025, Hu et al., 26 Mar 2026).

Taken together, these literatures suggest that zero-shot exploration is converging on a common principle: transferable performance on unseen targets depends less on generic novelty seeking than on exploration that is structurally aligned with the representation used at deployment. In semantic recognition that structure is the descriptor space; in multimodal reasoning it is the prompt-induced attention pattern; in reward-free RL it is the uncertainty geometry of the transferable value representation; and in embodied navigation it is the coupling among semantic priors, geometry, and persistent memory.