LIBERO-Long: Long-Horizon Robot Tasks
- LIBERO-Long is a suite of 10 long-horizon household- robot manipulation tasks that test transfer learning of declarative, procedural, and combined knowledge across tasks.
- Each task in LIBERO-Long involves diverse scenes, object interactions, spatial relations, and motor procedures, requiring the robot to perform multiple sub-goals sequentially.
- The principal metrics for LIBERO-Long include forward transfer (FWT), negative backward transfer (NBT), and area under the success-rate curve (AUC), which measure learning speed, task retention, and overall performance.
LIBERO-Long is a long-horizon robot-manipulation evaluation suite within LIBERO-100, designed to assess lifelong learning when declarative knowledge, procedural knowledge, and their combinations are entangled across tasks. It consists of exactly 10 long-horizon household-manipulation tasks. Unlike LIBERO-Spatial, LIBERO-Object, and LIBERO-Goal, which primarily isolate spatial relationships, object identities, or goals and motions, LIBERO-Long jointly varies objects, scenes, spatial arrangements, goals, and motor procedures. LIBERO-Long is used as the principal downstream evaluation suite after pretraining on the 90 short-horizon tasks of LIBERO-90 (Liu et al., 2023).
1. Definition and position within LIBERO
LIBERO as a whole contains 130 procedurally generated robot-manipulation tasks divided into four suites:
| Suite | Tasks | Principal variation |
|---|---|---|
| LIBERO-Spatial | 10 | Spatial relationships and object locations |
| LIBERO-Object | 10 | Object identities and types |
| LIBERO-Goal | 10 | Goals and motions |
| LIBERO-100 | 100 | Combined task collection |
LIBERO-100 is divided into LIBERO-90 and LIBERO-Long. LIBERO-90 contains 90 short-horizon tasks used as a pretraining and data source, whereas LIBERO-Long contains 10 long-horizon tasks used for downstream lifelong-learning evaluation. The ten LIBERO-Long tasks form an ordered continual-learning stream rather than a conventional supervised-learning dataset with an independently specified train, validation, and test partition.
The suite is intended to test transfer across three knowledge categories. Declarative knowledge concerns the objects and entities present, their visual appearance, and their locations. Procedural knowledge concerns how objects are manipulated, including reaching, grasping, moving, placing, opening, closing, stacking, and multi-step sequencing. LIBERO-Long combines these categories in tasks involving diverse scenes, object interactions, spatial relations, and motor procedures.
Examples explicitly associated with the suite include:
- “Put both the alphabet soup and the tomato sauce in the basket.”
- “Put the white mug on the left plate and put the yellow and white mug on the right plate.”
- “Put the yellow and white mug in the microwave and close it.”
These tasks involve multiple objects, relational placement, containers, appliances, and sequential subgoals. The supplied benchmark description does not provide a textual enumeration of all ten task names or a task-by-task inventory of their objects, scenes, goals, and motor primitives.
2. Formal lifelong-learning problem
Each task is represented as
where indexes the task in the lifelong sequence, is its task-specific initial-state distribution, and is its goal predicate. The robot uses one task-conditioned policy over state space and action space , with transition function and finite horizon .
After learning through task , the lifelong objective averages performance over all tasks encountered so far:
0
During learning of task 1, unrestricted access to previous tasks and their complete demonstration datasets is not assumed. This restriction creates the possibility of catastrophic forgetting. Because sparse-reward reinforcement learning is difficult in the benchmark setting, the experiments use lifelong imitation learning. A task dataset is
2
with trajectories of the form
3
Behavioral cloning minimizes a supervised action-prediction loss over demonstrations from the encountered tasks. Observations may be non-Markovian, so policies receive an observation history 4 rather than only the current observation. The policy outputs continuous end-effector actions through a Gaussian-mixture-model action head.
The benchmark evaluates success-rate curves rather than behavioral-cloning loss because small action errors can compound during execution. The principal lifelong metrics are forward transfer (FWT), negative backward transfer (NBT), and area under the success-rate curve (AUC). FWT measures the speed or strength of learning new tasks; lower NBT indicates less degradation on earlier tasks; AUC combines current-task learning with retention of previously learned tasks.
3. Procedural generation and task structure
LIBERO uses an extensible procedural-generation pipeline that can, in principle, generate infinitely many tasks. The pipeline composes five elements:
- A natural-language behavioral template.
- A set of simulator objects.
- A scene or fixture configuration.
- Object placements and initial states.
- A propositional or PDDL goal.
Behavioral templates are extracted from language descriptions in the Ego4D human-activity dataset and summarized into reusable forms. Simulator objects are inserted into the templates to produce manipulation instructions such as “Open the drawer of the cabinet.”
A scene layout is selected to match the instruction. For example, a kitchen scene is selected for “Open the top drawer of the cabinet and put the bowl in it.” The initial state 5 is specified in PDDL and includes object categories, fixtures, placement regions, object status, and spatial relationships.
Goals are expressed as conjunctions of predicates. Unary predicates include 6 and 7; binary predicates include 8 and 9. For the cabinet example, the goal is effectively
0
Simulation terminates when all goal predicates are true.
The controlled suites factorize knowledge types more cleanly. LIBERO-Spatial varies spatial relationships, LIBERO-Object varies object types, and LIBERO-Goal varies goals and motions while keeping objects and spatial relations fixed. LIBERO-Long instead contains entangled combinations of object interactions and versatile motor skills. The benchmark does not prescribe a canonical curriculum order for its ten tasks; experiments explicitly vary task order to measure order sensitivity.
4. Demonstrations, observations, and baseline policies
Every generated task, including each LIBERO-Long task, is supplied with 50 human-teleoperated trajectories collected using a 3Dconnexion SpaceMouse. This yields 500 demonstrations across the ten LIBERO-Long tasks. The experimental setup trains each task for 50 epochs using these demonstrations.
Evaluation uses simulator rollouts rather than additional human demonstrations: 20 test trajectories per task, each with a maximum length of 600. The observation includes perceptual information such as images, robot joint information, and gripper information. The supplied benchmark description does not specify the exact image resolution, number of camera views, joint-state dimensionality, action dimensionality, Gaussian-mixture component count, or trajectory-length distribution.
The original LIBERO study evaluates three vision-language policy architectures:
- ResNet-RNN: a ResNet visual encoder, LSTM temporal encoder, and FiLM-based language integration.
- ResNet-T: a ResNet visual encoder and transformer temporal decoder, with language represented as an additional token.
- ViT-T: a Vision Transformer visual encoder and transformer temporal decoder, with language used in both visual and temporal processing.
The architectures are matched at approximately 13.5 GFLOPs and use pretrained language embeddings by default, specifically BERT embeddings. Training uses Adam, batch size 32, a cosine learning-rate schedule from 1 to 2, 50 epochs per task, checkpoint evaluation every five epochs, and three random seeds: 100, 200, and 300.
The lifelong-learning protocols are:
- SeqL: sequential fine-tuning.
- ER: experience replay with up to 1,000 stored trajectories and 32 replay trajectories sampled per iteration.
- EWC: elastic weight consolidation using diagonal Fisher-information estimates, 3, and 4.
- PackNet: pruning, fine-tuning, and freezing, with 25% of relevant parameters selected per task.
- MTL: multitask learning with simultaneous access to all tasks, used as an approximate upper-bound reference rather than a lifelong protocol.
5. Reported lifelong-learning results
With ResNet-T fixed, the principal LIBERO-Long results are:
| 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 |
| MTL | — | — | 0.48 ± 0.01 |
Sequential fine-tuning has the highest FWT despite being the simplest protocol. ER provides the best AUC among the sequential lifelong methods, indicating the strongest overall balance between new-task learning and retention. PackNet provides the best retention, with NBT of 0.08, but has lower FWT and AUC than ER. EWC performs particularly poorly in AUC, with 0.02. MTL reaches 0.48 AUC, but has access to all tasks simultaneously and is therefore not directly comparable as a lifelong-learning method.
Architecture and lifelong algorithm interact substantially:
| Algorithm | Architecture | FWT | NBT | AUC |
|---|---|---|---|---|
| ER | ResNet-RNN | 0.16 ± 0.02 | 0.16 ± 0.02 | 0.08 ± 0.01 |
| ER | ResNet-T | 0.48 ± 0.02 | 0.32 ± 0.04 | 0.32 ± 0.01 |
| ER | ViT-T | 0.38 ± 0.05 | 0.29 ± 0.06 | 0.25 ± 0.02 |
| PackNet | ResNet-RNN | 0.13 ± 0.00 | 0.21 ± 0.01 | 0.03 ± 0.00 |
| PackNet | ResNet-T | 0.22 ± 0.01 | 0.08 ± 0.01 | 0.25 ± 0.00 |
| PackNet | ViT-T | 0.36 ± 0.01 | 0.14 ± 0.01 | 0.34 ± 0.01 |
Transformer temporal processing outperforms the recurrent ResNet-RNN architecture in this setting. Under ER, ResNet-T has the best AUC, 0.32. Under PackNet, ViT-T has the best AUC, 0.34, and the strongest FWT among PackNet configurations. These results indicate that architecture quality depends on the lifelong algorithm; no single visual encoder is uniformly best.
The language-embedding ablation uses ER and ResNet-T:
| Embedding | FWT | NBT | AUC |
|---|---|---|---|
| BERT | 0.48 ± 0.02 | 0.32 ± 0.04 | 0.32 ± 0.01 |
| CLIP | 0.52 ± 0.00 | 0.34 ± 0.01 | 0.35 ± 0.01 |
| GPT-2 | 0.46 ± 0.01 | 0.34 ± 0.02 | 0.30 ± 0.01 |
| Task-ID | 0.50 ± 0.01 | 0.37 ± 0.01 | 0.33 ± 0.01 |
The paper reports no statistically significant differences among these embeddings. The competitiveness of the Task-ID representation suggests that the embeddings may primarily distinguish tasks rather than supply extensive compositional semantic information.
6. Task ordering, pretraining, and later evaluations
LIBERO-Long is sensitive to task order. A study using ResNet-T with ER and PackNet evaluates five different orderings of the ten tasks. Performance varies substantially with ordering, and the difference is statistically significant for PackNet. The same algorithm can therefore behave differently depending on the sequence in which tasks are encountered. Attention analyses show that learning the same task at different positions produces different attention patterns, while later tasks tend to produce more diffuse attention.
Pretraining uses LIBERO-90, the 90 short-horizon tasks in LIBERO-100. The procedure is:
- Behavioral-cloning pretraining on LIBERO-90 for 50 epochs.
- Checkpoint saving every five epochs.
- Selection of the best checkpoint for each architecture.
- Initialization of downstream lifelong learning on LIBERO-Long.
- Comparison with models trained from scratch.
Naive supervised pretraining can hurt downstream LIBERO-Long lifelong-learning performance. Pretrained policies more often attend to task-irrelevant regions, and apparently more meaningful attention does not necessarily correspond to better performance. Consequently, attention maps are useful for diagnosis but are not reliable performance proxies.
Subsequent work has evaluated LIBERO-Long under additional long-horizon paradigms. These results are not directly interchangeable because they use different policies, protocols, demonstrations, task splits, or metrics.
- Static-dynamic visual token reuse: SD-VLA reports 92.4% aggregate LIBERO-Long success using OpenVLA-OFT, compared with 90.7% for the base model, while reducing latency from 741 ms to 437 ms and reporting a 1.70× speedup. Its dedicated temporal-memory evaluation is LIBERO-Memory, not LIBERO-Long (Qiu et al., 3 Feb 2026).
- Long-tail imitation learning: the related LIBERO-Core-LT construction uses ten selected LIBERO tasks with three head tasks and seven tail tasks. The dataset contains 154 demonstrations under a Pareto-skewed distribution. Approaching-Phase Augmentation improves success from 26.5% to 36.1%, but this is not the standard LIBERO-Long suite (Urieli, 30 May 2026).
- Reinforcement fine-tuning: LifeLong-RFT evaluates the standard ten-task LIBERO-Long suite with a six-task base phase followed by four incremental tasks. With 10 demonstrations for each new task and five replay demonstrations per previous task, it improves FWT from 61.0 to 74.2, NBT from 17.3 to 12.8, and AUC from 47.3 to 64.5 relative to NORA-Long SFT (Liu et al., 11 Feb 2026).
- Execution-time memory and recovery: HELM obtains 81.5% task success on LIBERO-Long with OpenVLA, compared with 58.4% for OpenVLA, by combining episodic memory, a state verifier, and rollback-based replanning. Increasing the history window to 32 yields 63.8%, indicating that longer recent context alone is insufficient (Zeng et al., 20 Apr 2026).
- Interleaved reasoning traces: IVLR reports 92.4% LIBERO-Long success using cached textual subgoals and visual keyframes. Removing the trace reduces success to 37.7%; text-only and vision-only traces reach 62.0% and 68.4%, respectively (Liu et al., 1 May 2026).
- VLA tools and hierarchical execution: VLAs-as-Tools improves 5 from 92.4% standalone SFT success to 97.2% with tool-aligned post-training and a VLA tool-family interface (Lei et al., 13 May 2026).
- Failure monitoring: Foresight evaluates failure detection rather than task completion. Its Transformer detector obtains ROC-AUC 6 and balanced accuracy 7 on LIBERO-Long using action-conditioned world-model latents and functional conformal prediction (Zhang et al., 22 Jun 2026).
- Smooth action representation: SMILE-SmolVLA improves LIBERO-Long success from 42.8% to 63.6%, while SMILE-Evo1 improves it from 89.8% to 98.2% through B-spline coefficient-space action generation. The 98.0% headline value for SMILE-Evo1 is its overall LIBERO score, not its Long-suite score (Park et al., 29 Aug 2026).
- Causal world-action modeling: Long-WAM reports 99.5% LIBERO-Long success for its inverse-dynamics-modeling configuration. Adding 2.4 seconds of visual history raises the robot-domain autoregressive model from 94.5% to 99.5%; removing future-video prediction reduces the corresponding result to 94.5% (Huang et al., 7 Oct 2026).
These later studies reinforce several properties exposed by LIBERO-Long: long-horizon execution depends on persistent task-state information, causal ordering, action smoothness, failure detection, recovery, and efficient use of demonstrations. They also demonstrate that reported success rates must be interpreted together with the evaluation protocol. LIBERO-Long does not have one universal result independent of policy architecture, task order, pretraining, action representation, rollout procedure, or intervention mechanism.