IPEC-COMMUNITY/language_table_lerobot Dataset Profile

IPEC-COMMUNITY/language_table_lerobot is a large-scale robotics dataset containing 442,226 episodes of xArm manipulator trajectories paired with language task specifications, totaling over 7 million frames captured at 10 frames per second. The dataset is formatted using the LeRobot v2.0 schema, storing each episode as a Parquet file with synchronized video recordings, and is designed for training language-conditioned manipulation policies such as vision-language-action models. It originates from the Language Table benchmark environment and has been structured for compatibility with the Open X-Embodiment ecosystem, enabling researchers and engineers to pretrain or fine-tune transformer-based robot policies on a substantial corpus of single-arm manipulation data.

The dataset is distributed under the Apache License 2.0, a permissive open-source license that explicitly permits commercial use, modification, and redistribution without requiring derivative works to be open-sourced. You may train proprietary models on this data, deploy those models in commercial robotic products, and retain full ownership of your trained weights and inference code. The license includes an express patent grant, reducing intellectual property risk, and requires only that you preserve copyright notices and include a copy of the license text with any distributions. Because Apache-2.0 does not impose copyleft obligations, it is particularly well-suited for enterprise procurement where legal teams need clear rights to incorporate open data into closed-source training pipelines and production systems.

This dataset is ideal for robotics teams building language-conditioned manipulation policies on xArm platforms, including researchers developing vision-language-action models, engineers fine-tuning teleoperation assistants, and machine learning teams pretraining large transformer policies for embodied AI. The scale—over 400,000 episodes—supports data-hungry architectures like RT-2 or Octo, and the LeRobot format ensures compatibility with modern imitation learning codebases. Teams working on instruction-following tasks, zero-shot generalization from linguistic commands, or cross-task transfer within tabletop manipulation will find the 127,605 distinct task specifications valuable for training models that generalize across diverse natural language prompts. The permissive license and high download count make it a low-risk choice for both academic benchmarking and commercial product development on xArm embodiments.

This dataset may not suit your needs if you are training policies for robots with significantly different morphologies, such as humanoid hands, mobile manipulators, or dual-arm systems, since it captures only xArm single-arm trajectories and transfer to other kinematic structures typically requires additional domain adaptation. If your application demands high-frequency control, the 10 FPS capture rate may lack the temporal resolution needed for contact-rich or high-speed tasks. The metadata does not confirm the presence of depth channels, force-torque sensors, or tactile feedback, so projects requiring multimodal sensory input beyond RGB video should verify modality coverage before integration. Additionally, without documented train-validation splits or episode success labels, teams needing curated subsets of high-quality demonstrations will need to implement their own filtering and quality-assurance steps, adding overhead to the data preparation pipeline.

The dataset employs the LeRobot v2.0 format, organizing episodes into 443 chunks of approximately 1,000 episodes each, with each episode stored as a Parquet file alongside a corresponding video recording. The Parquet schema captures robot state observations, action sequences, and metadata in columnar format, enabling efficient parallel reads and low-latency streaming during distributed training. Video files are synchronized frame-by-frame with trajectory data at 10 FPS, and the directory structure follows a predictable pattern that simplifies dataset sharding and episode indexing. Metadata in meta/info.json specifies the robot type, total frame count, chunk size, and split definitions, allowing loaders to preallocate buffers and schedule data workers appropriately. This structure is compatible with RLDS tooling and standard PyTorch or TensorFlow data pipelines, minimizing integration friction for teams already using LeRobot or Open X-Embodiment datasets.

Before procurement, verify that your compute infrastructure can handle the storage footprint of 7 million frames plus accompanying video, and confirm that your data loaders support Parquet and the LeRobot schema to avoid costly conversion work. Check whether your policy architecture requires modalities not explicitly listed, such as depth maps or proprioceptive force readings, since the metadata does not enumerate all sensor channels. Budget time to inspect a sample of episodes for task distribution, success rates, and annotation quality, as the dataset description truncates before providing these details. If you plan to combine this dataset with other embodiments for cross-robot pretraining, ensure your training harness can handle heterogeneous action spaces and observation dimensions. Finally, coordinate with your legal team to document the Apache-2.0 license in your data provenance records, confirm that no additional third-party assets introduce conflicting terms, and establish a process for preserving required copyright notices in your deployment artifacts.