IPEC-COMMUNITY/language_table_lerobot Dataset Profile

This is a large-scale robotics manipulation dataset containing 442,226 episodes of xArm robot trajectories performing language-conditioned tabletop tasks, totaling over 7 million frames. Each episode pairs natural language instructions with corresponding observation sequences and action trajectories recorded at 10 frames per second, covering 127,605 distinct manipulation objectives derived from the Language Table benchmark. The dataset ships in LeRobot v2.0 format with synchronized video recordings and is structured as 443 Parquet chunks for efficient training pipeline integration. Originally converted from the OpenX RLDS repository, the data captures pick-and-place, sorting, and spatial rearrangement tasks where language instructions specify target object configurations. All episodes represent successful task completions on an xArm manipulator platform, making the dataset suitable for supervised learning of vision-language-action policies, inverse dynamics modeling, and world-model pretraining for tabletop manipulation domains.

The dataset is released under the Apache License 2.0, which explicitly permits commercial use, modification, distribution, and private deployment without royalty payments. Robotics companies can train proprietary VLA models on this data, integrate resulting policies into commercial products, and deploy manipulation systems in revenue-generating applications provided they retain the Apache-2.0 license notice in any redistributed portions of the dataset. There are no field-of-use restrictions or publication requirements attached to the license. The permissive terms cover both trajectory data and video recordings, meaning visual representations learned from the 442,226 episodes carry no downstream licensing encumbrances. Teams building multi-dataset training pipelines should verify compatibility of other data sources, but this IPEC-COMMUNITY release itself imposes no constraints on model commercialization, cloud inference services, or embedded deployment in robotic hardware products.

Teams developing vision-language-action models for tabletop manipulation with language-conditioned task specifications will find the highest immediate value, particularly those targeting warehouse sorting, kitting operations, collaborative assembly prep, or household decluttering scenarios. The 127,605 language-task pairs provide substantial instruction diversity for training language encoders and grounding modules, while the xArm trajectories offer realistic action distributions for 6-DOF arm control. Startups and research labs building on the LeRobot ecosystem benefit from native format compatibility and established data loading utilities. Organizations requiring large-scale pretraining data for foundation models in manipulation can leverage the 7 million frames to learn visual representations, dynamics models, or inverse kinematics priors before fine-tuning on proprietary task distributions. The Apache-2.0 license and 919,489 download count indicate production-readiness and community validation, reducing procurement risk compared to bespoke data collection for initial VLA prototyping phases.

Teams building systems for contact-rich manipulation, deformable object handling, mobile manipulation, or dual-arm coordination should look elsewhere, as the Language Table scenarios focus exclusively on rigid-body tabletop tasks with fixed-base single-arm kinematics. The dataset lacks force-torque observations, tactile feedback, and high-frequency control signals necessary for precise insertion, assembly, or fragile object manipulation. Applications requiring bin-picking from clutter, human-robot handover, or navigation-manipulation integration will find limited transfer due to the constrained camera viewpoints and workspace geometry. Projects demanding failure recovery data, safety-critical edge case coverage, or adversarial robustness should recognize that episodes represent successful completions without explicit annotations for recoveries, collisions, or out-of-distribution handling. The 10 FPS sampling rate may prove insufficient for dynamic catching, high-speed pick-and-place, or reactive replanning scenarios requiring sub-100ms control loops. Finally, teams deploying on manipulator platforms with significantly different kinematic structures from the xArm will need additional domain adaptation or supplemental data to bridge action space gaps.

The LeRobot v2.0 format enables direct ingestion via the official LeRobot library's dataset loaders, which handle Parquet deserialization, video decoding, and trajectory normalization out of the box. Teams can load episodes by chunk index for distributed training or stream the full 442,226-episode corpus with the library's built-in samplers, leveraging the 1,000-episode chunk granularity for efficient multi-GPU data parallelism. For custom pipelines, the Parquet files are readable via standard Arrow or Pandas tools, with episode structure documented in the meta/info.json manifest file included in the repository. Observation and action space alignment requires validating that your target robot's state representation matches the xArm's joint configuration and end-effector pose encoding. The 10 FPS capture rate aligns with many commercial manipulator control frequencies, but teams running higher-rate policies should implement temporal interpolation or train action chunking modules. When mixing with other datasets for multi-source pretraining, ensure consistent normalization schemes for pixel observations, language embeddings, and action vectors, using the provided splits metadata to maintain clean train-validation boundaries across the 442,226 episodes.

Verify that your training infrastructure can handle the full dataset size of 7,045,476 frames distributed across 443 Parquet chunks, with each episode averaging roughly 16 frames at 10 FPS. Storage requirements include both trajectory data and 442,226 video files, so provision sufficient disk capacity and network bandwidth if streaming from HuggingFace Hub during training. The xArm action space consists of joint velocities or end-effector deltas depending on the control mode encoded in the trajectories, requiring inspection of sample episodes to confirm compatibility with your policy architecture's output head. The language task distribution spans 127,605 unique instructions, but teams should analyze the semantic diversity and linguistic complexity relative to their deployment domain's natural language interface requirements. Camera viewpoints in Language Table are predominantly overhead and fixed, so applications requiring viewpoint invariance or ego-centric perception may need augmentation strategies. Finally, confirm that the Apache-2.0 license terms satisfy your organization's legal and compliance requirements for commercial model training and deployment before integrating the dataset into production pipelines.