Meta-World: Benchmark and Applications

• Meta-World in reinforcement learning is a benchmark consisting of 50 simulated robotic manipulation tasks using a Sawyer robotic arm, designed to evaluate multi-task (MTRL) and meta-reinforcement learning (meta-RL).
• The benchmark aims to address task transfer and generalization, with tasks that include reaching, grasping, assembly, and tool use, varying across object configurations and manipulation primitives.
• In systems research, Meta-World broadly refers to a persistent, shared, interactive virtual environment that allows for the integration of digital twins, real-time communication, and digital economies.

Meta-World is a term with several technically distinct uses. In reinforcement learning, it denotes an open-source benchmark of 50 Sawyer robotic-arm manipulation tasks designed to evaluate multi-task reinforcement learning (MTRL) and meta-reinforcement learning (meta-RL). In systems and networking research, “Meta-World” denotes a persistent, shared, interactive virtual environment connecting physical and virtual worlds. More recent work uses MetaWorld as the name of hierarchical world-model, offline meta-RL, and multi-agent video-generation frameworks. These uses share an emphasis on generalization across tasks, agents, environments, or viewpoints, but they are not a single unified scientific object.

1. Terminology and conceptual scope

The original Meta-World benchmark was introduced to address the narrow task distributions used in earlier meta-RL research. Such distributions often varied only a parameter, such as a target location or running velocity, and therefore tested adaptation to parametric variation rather than acquisition of qualitatively different behaviors. Meta-World combines non-parametric variation across manipulation skills with parametric variation in object and goal configurations (Yu et al., 2019).

The benchmark’s central question is whether experience with multiple related manipulation problems can accelerate acquisition of another task. Its task family includes reaching, pushing, pulling, grasping, placing, articulation, tool use, insertion, assembly, and object retrieval. The benchmark preserves common structure through a shared Sawyer arm, tabletop workspace, MuJoCo physics, low-dimensional observation and action interfaces, and reusable manipulation primitives, while varying object affordances, mechanisms, reward functions, and action sequences.

A separate body of research uses “Metaverse” to denote a candidate next-generation Internet rather than a robotics benchmark. This conception describes a shared and persistent virtual space connecting multiple three-dimensional environments through the Internet, with users represented by avatars. Its broader formulation integrates social interaction, immersive interfaces, digital twins, real-time communication, virtual economies, distributed systems, and physical-world sensing and actuation. There is no universally accepted definition of the Metaverse (Cheng et al., 2022).

The name has also been reused for methods that address related generalization problems:

• “Meta-World Conditional Neural Processes” uses a conditional probabilistic transition model to generate synthetic adaptation experience, although its experiments use a 2D point-agent environment, cartpole, and Walker-2D rather than the canonical robotic Meta-World suite (Ada et al., 2023).
• MAMBA is a recurrent latent world-model approach to meta-RL, evaluated on Point Robot Navigation, Escape Room, Rooms-$N$, Reacher-$N$, Visual Reacher, Panda Reacher, Humanoid-Dir, and related environments rather than Meta-World (Rimon et al., 2024).
• Meta-DT applies offline conditional sequence modeling to three Meta-World tasks—Reach, Sweep, and Door-Lock—with hidden object or goal configurations (Wang et al., 2024).
• “MetaWorld: Hierarchical Semantic Planning, Expert Transfer, and Latent Physical Control” addresses humanoid loco-manipulation through VLM-driven skill selection, expert-policy fusion, and latent MPC on Humanoid-Bench (Shen et al., 24 Jan 2026).
• “MetaWorld: Scaling Multi-Agent Video World Model from Single-view Video Data” addresses synchronized multi-agent video generation using monocular geometric decomposition, identity conditioning, and cross-view latent alignment (Hu et al., 1 Jun 2026).

Thus, the benchmark and the virtual-world concept are related primarily by the general problem of constructing systems that transfer structure across heterogeneous tasks or worlds. They should not be treated as interchangeable terms.

2. The robotic manipulation benchmark

Meta-World contains 50 simulated manipulation environments using a Sawyer robotic arm, a common tabletop workspace, MuJoCo contact dynamics, and shared low-dimensional interfaces. The tasks include turn-on-faucet, sweep, assemble-nut, push, pull-lever, turn-dial, push-with-stick, basketball, hammer, slide-plate, button pressing, soccer, drawer and door manipulation, peg insertion, pick-and-place, window manipulation, lock and unlock operations, and bin picking (Yu et al., 2019).

The task suite combines two kinds of variation:

• Parametric variation: object and goal positions are randomized within a task family.
• Non-parametric variation: different task families require qualitatively different objects, mechanisms, interaction modes, and sometimes ordered combinations of primitives.

A task is described as

$$T=(\text{reward function},\text{initial object position},\text{target position}).$$

Some tasks move one object toward a variable goal; others manipulate two objects toward a fixed configuration. Complex behaviors may require ordered sequences such as grasping a stick, positioning it near another object, and using it to push or pull a second object.

The action is a four-dimensional continuous vector consisting of three end-effector displacement components and gripper control:

$$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$

The observation is 39-dimensional. It contains the current and previous 18-dimensional observations together with three goal coordinates. The observation components include end-effector position, gripper state, positions and quaternion orientations of up to two objects, and the goal position. When a second object or a goal is absent, the corresponding coordinates are zeroed.

The benchmark is primarily state-based rather than visual. Object positions and orientations are exposed directly, so the initial benchmark does not require image-based perception. This makes it suitable for studying task transfer, representation sharing, task inference, and optimization independently of visual recognition.

Reward design and evaluation

The original benchmark used shaped rewards intended to make individual tasks solvable, normalize reward scales approximately between $0$ and $10$, expose common manipulation structure, and reward intermediate behavior such as reaching, grasping, caging, lifting, and placing. Smooth tolerance functions and Hamacher products combine geometric and behavioral constraints.

Success rather than shaped reward is the official evaluation criterion. Task-specific success thresholds include approximately $5$ cm for reaching and pushing, $7$ cm for pick-and-place and faucet manipulation, $8$ cm for doors and basketball, $3$ cm for drawer opening, $N$0 cm for button pressing, and $N$1 cm plus a grasp condition for stick manipulation. Other tasks use joint-angle, insertion, nail-travel, or assembly criteria.

The benchmark commonly uses a 500-step episode horizon and discount factor $N$2. Positions may be randomized for generalization protocols or fixed for protocols intended to isolate skill acquisition.

3. Multi-task and meta-reinforcement-learning protocols

Meta-World distinguishes simultaneous multi-task learning from adaptation to unseen tasks. In MTRL, a task-conditioned policy receives a task descriptor, usually a one-hot task identifier:

$N$3

The task identity and goal are available to the agent. Evaluation uses the same task collection used for training, although configurations may vary depending on the protocol.

In meta-RL, the agent trains on one task distribution and adapts to held-out tasks. Task identity and goal information are not directly supplied. The agent must infer task-relevant information from interaction history and use a small amount of experience to improve behavior on a new task.

The principal settings are:

• ML1: adaptation across 50 training and 50 held-out configurations within a single task family, including reaching, pushing, and pick-and-place. Goal positions are withheld from observations.
• MT1: learning across 50 configurations of one task family with goal positions provided.
• MT10: one task-conditioned policy trained on 10 environments and 50 configurations per environment, producing 500 training task instances. Positions are fixed.
• MT50: training on all 50 environments and 2,500 task instances, generally with task IDs and fixed positions.
• ML10: training on 10 environments and adaptation to five held-out environments, with randomized positions and no task IDs.
• ML45: training on 45 environments and adaptation to five held-out environments, also with randomized positions and no task IDs.

The hierarchy moves from configuration adaptation to simultaneous learning across distinct skills and finally to held-out-task generalization. This ordering is conceptual rather than a strict difficulty ranking, because the protocols emphasize different sources of difficulty.

Baseline results

Seven methods were evaluated in the original study: multi-task PPO, multi-task TRPO, multi-task SAC, Task Embeddings, RL$N$4, MAML, and PEARL. On MT10 and MT50, the reported average maximum success rates were:

Method	MT10	MT50
Multi-task PPO	30.5%	35.4%
Multi-task TRPO	31.3%	21.0%
Task Embeddings	20.9%	11.8%
Multi-task SAC	68.3%	38.5%

Multi-task SAC was substantially more sample-efficient than the on-policy multi-task methods and achieved the highest MT10 and MT50 values in the reported table. Scaling from 10 to 50 task families sharply reduced performance.

On ML10, RL$N$5 achieved $N$6 meta-training success but only $N$7 on held-out tasks. MAML achieved $N$8 on training tasks and $N$9 on held-out tasks, while PEARL achieved $$T=(\text{reward function},\text{initial object position},\text{target position}).$$0 and $$T=(\text{reward function},\text{initial object position},\text{target position}).$$1, respectively. On ML45, MAML obtained the highest reported held-out success at $$T=(\text{reward function},\text{initial object position},\text{target position}).$$2, compared with $$T=(\text{reward function},\text{initial object position},\text{target position}).$$3 for RL$$T=(\text{reward function},\text{initial object position},\text{target position}).$$4 and $$T=(\text{reward function},\text{initial object position},\text{target position}).$$5 for PEARL.

These results establish that individual task solvability does not imply simultaneous multi-task competence or held-out-task adaptation. SAC can learn all 50 tasks to some degree, and PPO can solve a large majority of them, yet multi-task and meta-RL performance remains modest.

4. Benchmark versioning and Meta-World+

The original benchmark underwent undocumented implementation and reward changes. In particular, the initial dense reward system, referred to as V1, was replaced in later usage by a different system, V2. Papers sometimes reported “Meta-World” results without identifying the release, reward version, task definition, or evaluation protocol, making numerical comparisons unreliable (McLean et al., 16 May 2025).

V1 rewards were based on structured pick-and-place-like sequences, including moving above an object, opening and closing the gripper, lifting, transporting, and lowering toward a goal. Some task-specific rewards could reach approximately $$T=(\text{reward function},\text{initial object position},\text{target position}).$$6 per step and could encode behavior not strictly necessary for task success.

V2 rewards were designed around fuzzy geometric constraints. Their values generally lie in a narrower range, approximately $$T=(\text{reward function},\text{initial object position},\text{target position}).$$7 to $$T=(\text{reward function},\text{initial object position},\text{target position}).$$8, after rescaling. V2 was intended to make PPO achieve high success within approximately 20 million steps, reduce unnecessary motion prescriptions, equalize reward scales across tasks, and maintain Markovian reward definitions.

The two reward systems are not merely rescaled copies. They induce different optimization problems and affect algorithms differently. In the reported MTRL experiments, all tested methods performed better with V2 than V1. For example, MT10 success for PaCo increased from $$T=(\text{reward function},\text{initial object position},\text{target position}).$$9 under V1 to $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$0 under V2, while MT50 success increased from $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$1 to $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$2. The paper attributes much of this sensitivity to shared value-function and Q-function estimation.

Meta-World+ preserves both V1 and V2 through an explicit API rather than introducing another reward version. It also provides:

• Gymnasium compatibility and modern MuJoCo Python bindings;
• explicit task-set selection;
• MT10, MT25, MT50, ML10, ML25, ML45, and custom configurations;
• separate training and testing environments for meta-RL;
• standardized evaluation helpers;
• reporting procedures using 10 seeds, interquartile means, and 95% confidence intervals in the stated experiments.

The benchmark’s principal metric is mean success rate over evaluation tasks. For task $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$3, if $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$4 of $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$5 evaluation episodes succeed,

$$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$6

and the aggregate score is the mean across tasks.

Meta-World+ does not change the core Sawyer morphology, 39-dimensional observation, four-dimensional action, or 50 task definitions according to the supplied material. Its main contribution is version awareness, reproducible interfaces, controllable task sets, and standardized evaluation. Researchers are expected to report the software version or commit, reward version, exact task suite, task names, train/test split, goal visibility, preprocessing, seeds, evaluation episodes, aggregation method, confidence intervals, and implementation details.

5. Algorithmic extensions for Meta-World-like generalization

Several later methods use Meta-World-like task distributions while addressing representation learning, model-based adaptation, or offline learning.

Conditional world-model adaptation

MW-CNP learns a conditional neural process for transition dynamics. A context set of observed transitions is encoded by shared parameters and aggregated by a permutation-invariant mean:

$$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$7

The decoder predicts a probabilistic next-state distribution conditioned on the query state-action pair and $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$8. At test time, a short rollout from an unseen environment produces a latent representation of its dynamics. The learned model then generates “hallucinated rollouts” that are combined with the real rollout for policy adaptation.

The experiments use a force-field point agent, cartpole with hidden sensor bias, and Walker-2D with randomized physical parameters. On the point-agent experiment, MW-CNP achieved a post-update loss of $$a_t=(\Delta x,\Delta y,\Delta z,g), \qquad a_t\in[-1,1]^4.$$9, compared with $0$0 for NORML using one real rollout and $0$1 for a 25-rollout NORML oracle. These results support the use of learned synthetic experience under shared task-family assumptions, but they do not establish performance on the canonical Meta-World manipulation suite (Ada et al., 2023).

Recurrent latent world models

MAMBA modifies DreamerV3 for meta-RL by encoding the complete meta-episode with a recurrent state-space model. The latent state receives observations, rewards, and time, allowing it to preserve information collected across sub-episodes. The method uses imagined latent rollouts for policy learning and schedules the world-model training horizon from short prefixes to the full meta-episode.

MAMBA was not evaluated on Meta-World. Its reported environments include Rooms-$0$2, Reacher-$0$3, Point Robot Navigation, Escape Room, Visual Reacher, Panda Reacher, and Humanoid-Dir. The results show improved returns and interaction efficiency on many of these tasks, including $0$4 on Reacher-3 and $0$5 on Reacher-4. The paper therefore provides evidence for full-context recurrent world models in decomposable meta-RL tasks, not direct evidence on MT10, MT50, ML10, or ML45 (Rimon et al., 2024).

Offline sequence modeling

Meta-DT combines a context-aware world model with a causal transformer. The context encoder uses a four-step history and produces a 16-dimensional task embedding. The world model predicts reward and transition consequences, after which it is frozen. The transformer receives the task representation, recent state-action history, returns-to-go, and a short complementary prompt selected from a trajectory segment with the largest world-model prediction error.

On three Meta-World task types—Reach, Sweep, and Door-Lock—the method uses 15 training tasks and five held-out test tasks. The reported metric is episodic return rather than success rate. Using Medium offline datasets, Meta-DT obtained:

Task	Meta-DT return
Reach	$0$6
Sweep	$0$7
Door-Lock	$0$8

Meta-DT outperformed Prompt-DT, Generalized DT, CORRO, and CSRO in the reported few-shot comparisons. The evidence is limited to three task types and does not establish broad performance across the complete Meta-World suite or complete zero-shot success-rate performance (Wang et al., 2024).

Contextual latent world models

Self-Predictive Contextual Offline Meta-RL, also described as a Contextual Latent World Model, learns task representations jointly with a discrete latent dynamics model. The context representation is the mean embedding of transition tuples:

$0$9

Observations are mapped to discrete latent codes through finite scalar quantization. The task-conditioned model predicts future latent codes and rewards over a five-step horizon. Its objective combines multi-step temporal consistency with an InfoNCE contrastive loss. IQL then learns a policy directly over latent codes and task representations rather than using the world model for explicit long-horizon planning.

The Meta-World experiments use 39-dimensional observations and four-dimensional actions, ML1, ML10, and ML45 protocols, and offline datasets generated by separate DroQ agents. Each task dataset contains approximately 1,000 trajectories collected across roughly $10$0 environment steps. The default context size is 64 transitions, the task representation dimension is 5, and evaluation uses task success rate with six random seeds, 95% confidence intervals, interquartile means, and optimality gaps.

The method generally outperforms FOCAL, CSRO, DORA, UNICORN-SS, and UNICORN-SUP on reported ML1 tasks. Examples include $10$1 on Button-press and Window-open, $10$2 on Door-open, $10$3 on Faucet-open, and $10$4 on Peg-insert-side. Assembly yields $10$5 for all methods because the offline dataset contains no successful trajectories. On ML10 and ML45, the method improves both training-task and held-out-task performance, although unseen-environment generalization remains difficult (Nakheai et al., 3 Mar 2026).

6. Broader virtual-world and infrastructure interpretations

In Metaverse research, the Meta-World concept is a persistent, shared, interactive environment linking physical and virtual worlds. It is expected to support social presence, avatars, real-time interaction, digital twins, interoperable identities and assets, scalability, accessibility, digital economies, and physical-world integration (Cheng et al., 2022).

The enabling technologies include XR, 5G, edge computing, blockchain, NFTs, machine learning, digital twins, computer vision, haptics, and brain-computer interfaces. Requirements include low latency, high throughput, state synchronization, large-scale rendering, storage, identity management, and governance. The paper reports that current social-VR systems remain proprietary, device-dependent, limited in user count, weakly interoperable, and based on simplified avatars.

Measurements of Workrooms and AltspaceVR illustrate the gap between current platforms and the broader vision. In a two-user experiment, Workrooms produced approximately $10$6–$10$7 Mbps downlink and $10$8 Mbps uplink per user, while AltspaceVR produced less than $10$9 Mbps post-loading throughput after downloading $5$0–$5$1 MB plus $5$2–$5$3 KB of event content. In five-user experiments, Workrooms downlink increased by approximately $5$4 Mbps per additional user and AltspaceVR by approximately $5$5 Mbps. These observations indicate approximately linear traffic growth in the tested server-centered architectures, but they do not predict the requirements of a future full-body, volumetric, or haptic system.

Slicing4Meta proposes an infrastructure architecture for Metaverse-as-a-Service. Its five layers are the Physical World Layer, Technology Layer, MaaS and Resource Integration Layer, Metaverse Service Instance Layer, and AI Layer (Liu et al., 2022). The architecture treats communications, computation, storage, sensing, devices, rendering, data, software, platforms, and technologies as coordinated resources rather than treating the network as an isolated communication pipe.

A Metaverse Service Instance represents a running service, while an MSI cluster groups instances of the same type. The paper distinguishes AR/VR immersive services from digital-twin services. Global AI-assisted controllers provide system-wide planning, while local controllers instantiate, scale, and monitor services. The virtual-travel case uses a total rendering capacity of $5$6, with $5$7 pixels and $5$8 allocated per virtual object. The reported qualitative finding is that, under fixed rendering capacity, users with higher data rates obtain better Meta-Immersion.

The Metaverse vision raises unresolved issues involving interoperability, privacy, identity, ownership, content moderation, harassment, governance, regulation, digital assets, decentralized economies, and physical-world safety. Slicing4Meta addresses resource and service orchestration but does not provide a complete social, legal, economic, or governance theory.

7. Recent world-model formulations named MetaWorld

The name has increasingly been applied to architectures that combine semantic reasoning, learned dynamics, expert transfer, or multi-agent observation generation.

Hierarchical humanoid control

The hierarchical MetaWorld framework separates a VLM-driven semantic layer from a physical control layer. The VLM maps an instruction and environment information to weights over a library of expert policies:

$5$9

These weights are fused with state-aware expert-selection probabilities:

$7$0

with reported $7$1. The resulting expert mixture supplies a reference action to a TD-MPC2-style latent controller. The controller uses a short MPC horizon $7$2 and an expert-guidance coefficient $7$3.

Experiments on Humanoid-Bench evaluate Stand, Walk, Run, and Door. MetaWorld reports returns of $7$4, $7$5, $7$6, and $7$7, respectively. The framework is compared with DreamerV3 and TD-MPC2 and reports faster convergence and higher returns, particularly for Run and Door. Removing the VLM reduces door-task performance to $7$8, removing dynamic selection yields $7$9, and removing expert guidance yields $8$0. The system remains dependent on the coverage and transferability of its expert library and has not been evaluated on physical hardware (Shen et al., 24 Jan 2026).

Multi-agent video world modeling

A different MetaWorld framework generates synchronized egocentric videos of multiple agents from single-view video data. It consists of Monocular World-State Unrolling, the Subject-Aware World Generator, and World-State Alignment (Hu et al., 1 Jun 2026).

MWSU estimates camera motion with MoGe-2, depth with Depth Anything, and subject masks with SAM 2. Masked image points are back-projected into 3D and represented using moving 3D Gaussian spheres. Global trajectories are transformed into each agent’s camera coordinates, allowing each generated branch to depict the same agents and environment from a different viewpoint.

The Subject-Aware World Generator uses a Phantom-Wan2.1 14B video Diffusion Transformer. RGB and depth trajectory conditions control subject location and geometry, while identity reference images preserve subject appearance. World-State Alignment inserts frame-restricted inter-branch cross-attention into every transformer layer. Two-agent training is extended to $8$1 agents through cyclic communication rather than dense $8$2 attention.

In the full-3D evaluation, MetaWorld achieved a cross-view consistency score of $8$3, compared with $8$4 for VerseCrafter and $8$5 for MultiWorld. Its trajectory consistency was $8$6, compared with $8$7 and $8$8, respectively. The system does not model active manipulation, collisions, contact dynamics, or causal physical interaction; its “physics-driven” behavior primarily concerns geometric and motion consistency.

Across these uses, Meta-World research converges on a broad methodological theme: systems should not merely memorize isolated tasks, configurations, trajectories, or views. They should infer latent structure that supports transfer across manipulation skills, hidden environments, virtual services, semantic instructions, or multiple observations of a shared world. The benchmark remains the most standardized use of the term, while the newer frameworks extend the concept toward model-based adaptation, offline generalization, hierarchical control, virtual-world infrastructure, and multi-agent world modeling.