LIBERO & Meta-World Benchmarks

• LIBERO and Meta-World are benchmarks defined as standardized platforms for evaluating multi-task, meta-reinforcement, and lifelong learning in robotic manipulation.
• LIBERO emphasizes lifelong knowledge transfer using procedural generation, detailed skill blending, and extensive human demonstration data.
• Meta-World focuses on fixed multi-task/meta-RL challenges with standardized reward scaling, rigorous reproducibility protocols, and controlled task variability.

LIBERO and Meta-World are benchmarks that rigorously advance the evaluation landscape for multi-task, meta-reinforcement learning, and lifelong learning in robotic manipulation. They aim to provide standardized, reproducible, and extensible platforms for algorithmic comparison, formalizing diverse task sets, reproducibility protocols, and deep experimentation infrastructure. While Meta-World (and its successor Meta-World+) focuses on standardized multi-task/meta-RL with articulated reward-scale and task-set controls, LIBERO specifically targets lifelong knowledge transfer, procedural/declarative skill blending, procedural generation, and task-ordered robustness, all with large-scale human demonstration data. Together, these benchmarks form the evaluative backbone of contemporary RL and continual learning for manipulation.

1. Structural Overview and Motivation

Meta-World comprises a suite of 50 manipulation tasks implemented in MuJoCo using a Sawyer arm, each representing a finite-horizon MDP sharing a 39-dimensional state and 4-dimensional action space. The environments incorporate broad parametric variability via randomization over object placements and goals, spanning reaching, pushing, object manipulation, and mechanism actuation (Yu et al., 2019). This design motivates the development of algorithms capable of generalizing across fundamentally distinct skills.

LIBERO targets lifelong learning in robot manipulation where agents are expected to accumulate and transfer both declarative (conceptual/object) and procedural (motor/behavioral) knowledge across a potentially unbounded task stream. Unlike fixed-task benchmarks, LIBERO uses procedural generation based on Robosuite and PDDL, systematically varying objects, spatial layouts, and behavioral goals. Four main task suites—LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-100—enable granular ablation along knowledge-type boundaries, supporting detailed investigations of transfer properties (Liu et al., 2023).

2. Task Design, Procedural Generation, and Suite Definition

Meta-World structures each task as a unique MDP:

• State: Cartesian gripper position, opening, object poses, and zero-padded goal location ($s\in\mathbb{R}^{39}$).
• Action: Cartesian displacement ($\Delta\mathbf{x}\in\mathbb{R}^3$) and normalized gripper torque ($a\in[-1,1]^4$).
• Task suite includes fixed primitives (e.g., Reach, Push, PegInsert, DoorOpen) and parametric variations, resulting in effectively infinite intra-task distributions (Yu et al., 2019).

Task sets—MT10, MT50 (multi-task), ML10, ML45 (meta-RL)—differ in train/test splits, adaptation requirements, and object/goal diversity. Meta-World+ further resolves historical ambiguities in reward scaling and task-set drift by versioning rewards ("v1" vs. "v2"), freezing all task sets, and supporting custom combinations via explicit IDs (McLean et al., 16 May 2025).

LIBERO operationalizes procedural generation as follows:

• Behavioral language templates (mined from Ego4D) instantiate natural-language tasks, grounded to simulator objects and scenes.
• Scenes sampled across kitchen, living room, and study, using PDDL “:regions” for flexible placement.
• Goal predicates form conjunctions of atomic/unary/binary relations (e.g., Open(X), In(A,B), On(A,B)) (Liu et al., 2023).

The key suites:

• LIBERO-Spatial: tests spatial relation transfer with identical objects in different positions.
• LIBERO-Object: tests object identity transfer with fixed behaviors.
• LIBERO-Goal: tests procedural transfer with diverse goals and constant objects/layout.
• LIBERO-100: combines all shifts in a mixed, large-scale format for real-world emulation. Altogether, 130 tasks with 6,500 human demonstration episodes are available.

3. Reproducibility, Evaluation Metrics, and Experimental Protocols

Meta-World+ enforces reproducibility via:

• Separate environment IDs for reward-function variants, detailed API changelogs, and broadcasted task-set definitions.
• Gymnasium compliance, enabling drop-in compatibility with RLlib, Stable Baselines3, and other Gymnasium-native frameworks.
• All evaluations use fixed random seeds, report interquartile means (IQM) with 95% confidence intervals, and evaluate over 10 seeds (McLean et al., 16 May 2025).
• Metrics:
  • Success rate per task $i$: $$SR_i = \frac{1}{N}\sum_{j=1}^N 1\{\mathrm{episode\ succeeds}\}$$.
  • Aggregate success: $$\overline{SR} = \frac{1}{|\mathcal{T}_\mathrm{eval}|} \sum_{i\in\mathcal{T}_\mathrm{eval}} SR_i$$.
  • Return: mean discounted sum, reported analogously.

LIBERO formalizes lifelong evaluation with:

• Forward Transfer (FWT): measures improvement on new tasks due to prior experience.
• Negative Backward Transfer (NBT): quantifies forgetting or interference on previous tasks post-learning.
• Area Under Curve (AUC): overall retained performance across the sequence. Formally,

$$FWT = \frac{1}{K}\sum_{k=1}^K \left[\frac{1}{|\mathrm{Epochs}|}\sum_e c_{k,k,e} - \mathrm{base\_rate}_k\right],\quad NBT = \frac{1}{K}\sum_{k=1}^K \frac{1}{K-k}\sum_{\tau=k+1}^K \left[c_{k,k} - c_{\tau,k}\right],\quad AUC = \frac{1}{K}\sum_{k=1}^K \frac{1}{K-k+1}\left(FWT_k + \sum_{\tau=k+1}^K c_{\tau,k}\right).$$

Agents are evaluated under multiple task orderings, enabling direct measurement of curriculum robustness (Liu et al., 2023).

4. Baseline Algorithms, Architectures, and Empirical Results

Meta-World’s evaluations span:

• Multi-task RL (PPO, TRPO, SAC) with one-hot or embedding task conditioning.
• Meta-RL (RL$^2$, MAML, PEARL), where context inference and fast adaptation are salient.
• Notable findings: Success rate on MT10 via SAC$_\mathrm{MT}$ is ~68%, dropping to ~38% on MT50; meta-test success on ML10 is $~32$–$36\%$ (RL$^2$, MAML), with severe performance drop as the number of tasks grows (Yu et al., 2019).
• "Ray interference" and inefficient context representation are primary impediments at scale.

LIBERO benchmarks:

• Visual architectures: ResNet-RNN, ResNet-Transformer (ResNet-T), ViT-Transformer (ViT-T) combining vision with language/task ID embeddings.
• Lifelong algorithms: Sequential Finetuning (SeqL), Experience Replay (ER), EWC (regularization), PackNet (dynamic-architecture), and Multi-Task Learning (upper bound) (Liu et al., 2023).
• Surprises:
  • SeqL, which does not explicitly prevent forgetting, outperforms sophisticated LLDM methods in forward transfer on all suites.
  • No single visual backbone is dominant: ResNet-T is optimal for ER, ViT-T for PackNet and object-based transfer.
  • Task-language embedding choice (BERT, CLIP, GPT-2, Task-ID) does not affect performance, indicating encoding operates as an index rather than informative semantic vectorization.
  • Naive pretraining on the LIBERO-90 subset consistently degrades downstream lifelong learning performance.

Selected empirical results (ResNet-T, LIBERO-Long):

Algorithm	FWT↑	NBT↓	AUC↑
SeqL	0.54 ±0.01	0.63 ±0.01	0.15 ±0.00
ER	0.48 ±0.02	0.32 ±0.04	0.32 ±0.01
EWC	0.13 ±0.02	0.22 ±0.03	0.02 ±0.00
PackNet	0.22 ±0.01	0.08 ±0.01	0.25 ±0.00

5. Technical Ergonomics, Benchmark Interoperability, and Extensions

Meta-World+ enhances usability and extensibility through:

• Single-line, vectorized environment instantiation with explicit control over reward versioning, task selection, and random seeds.
• YAML/JSON config files for transparent benchmarking, supporting reproducible baselines.
• All historical reward/task choices preserved; cross-benchmark comparison enabled by Gymnasium registration, shared evaluation code (IQM, CI), and vectorized runtime (McLean et al., 16 May 2025).

LIBERO exposes high-quality demonstration data as 6,500 human teleoperation trajectories, modular procedural generation, and combinable task suites. The design enables curriculum construction and ablation across knowledge dimensions. Tasks and demos are readily extendible by integrating new object libraries, templates, and scene types (Liu et al., 2023).

Meta-World+ and LIBERO can be integrated:

• Pretraining on LIBERO’s procedurally generated suites followed by lifelong adaptation on Meta-World’s fixed tasks.
• Cross-benchmark curricula or ordering-sensitivity analyses to probe robustness.
• Combining Meta-World’s diverse robotic arms and structured task sets with LIBERO’s generative pipeline enables scaling to 200+ manipulation environments.

6. Open Research Questions and Future Directions

Both benchmarks expose pressing challenges:

• Learning algorithms with improved forward and backward transfer tradeoffs (e.g., combining replay and regularization).
• Policy architectures that factor and disentangle spatial, temporal, and task-language cues more efficiently.
• Pretraining objectives that yield robust lifelong transfer, possibly with contrastive or language-grounded components.
• Task order robustness: PackNet demonstrates high variance to task permutation, indicating curriculum design remains unsolved.
• Privacy-preserving protocols for lifelong imitation learning with human demonstration data (LIBERO).

Each benchmark explicitly recognizes the limitations of current RL/LDM methods, particularly in the face of real-world complexity, reward-scale inconsistency, and unbounded or entangled task distributions (Liu et al., 2023, McLean et al., 16 May 2025, Yu et al., 2019). Their extensible frameworks provide a foundation for algorithmic research that is directly comparable and rigorously repeatable.

7. Comparative Positioning and Significance

Meta-World+ represents a rigorously versioned, formally specified multi-task and meta-RL platform ensuring reward-scale consistency, reproducible baselines, and ergonomic extensibility for Sawyer-based manipulation (McLean et al., 16 May 2025). LIBERO shifts the field toward scalable, lifelong knowledge transfer, with explicit ablations along declarative and procedural axes, and the procedural generation of open-ended manipulation challenges (Liu et al., 2023). Combined, these suites constitute the primary testbeds for developing and comparing robust, generalist, and adaptable robotic manipulation agents under both bounded and unbounded task regimes.