Open datasets, benchmarks and evaluation suites for robot learning — what was released, under which licence, and how big.
EgoSteer-RealWorld, the real-robot dataset used in EgoSteer, holds 54,454 teleoperated episodes (192 hours, 20.75 M frames) of bimanual dexterous manipulation across 193 tasks on a RealMan dual-arm robot, according to its Hugging Face dataset card.
FoldNet++ spans 6 robotic embodiments, 1K T-shirts, 1K environmental assets and 120K episodes with rich annotations. It generates physically simulatable T-shirts with annotated semantic keypoints, from which a unified rule-based framework produces demonstrations for each embodiment. The authors report that visuomotor policies trained solely on this synthetic data achieve over 90% end-to-end task success when deployed directly to unseen real-world environments and previously unseen T-shirts.
GE-Act 2.0 trains every parameter from random initialisation on embodied manipulation data, rather than adding action prediction to a pretrained video generator. AgiBot reports that scaling the corpus from 300 to 30,000 hours lifted zero-shot success on its G1-OP evaluation from 17.1% to 44.1%, and the number of tasks with any success from 39 to 76, with towel-folding and pen-recapping appearing only at the larger scales. It says performance had not saturated at 30,000 hours.
MEMOBench is a process-level memory benchmark for robotic manipulation, testing policies on information that has left the frame.
The authors release 1,500 hours of bimanual household demonstrations and train XR-2, a vision-language-action model, on them, reporting that success rises steadily with more demonstration data and with DAgger corrections from live human interventions.
AGIBOT has open-sourced 11,430 real-world trajectories across 14 industrial and household tasks, including 1,024 successful and 1,369 failed policy rollouts. Annotations cover task progress, completion, errors, disturbances and human interventions, giving structured execution feedback rather than demonstrations alone. It is the reinforcement-learning theme of AGIBOT WORLD 2026, an open-source dataset initiative organised into themes, and the data spans expert demonstrations, autonomous deployment and human-in-the-loop correction.
embodied-sync is an open-source Python library for aligning and validating multi-sensor robot datasets, handling different sample rates, clock skew and drift, late packets and missing frames. It fits around UMI, LeRobot, ROS 2 and MCAP, and the author describes the project as alpha. The pitch is that cameras run at one rate and robot state at another, packets arrive late, and a device reconnect can silently reset its clock, so a policy trains on observations that never lined up.
Hebbian Robotics, part of Y Combinator's summer 2026 batch, has open-sourced HFlow, an SDK for multimodal data pipelines in robotics. It checks recordings for frozen cameras, drifting streams, missing topics and duplicates, runs transformations and labels, and stamps each processed episode with its provenance in a queryable catalogue, with MCAP as its input and output format.
ViFailback, the dataset from a CVPR 2026 paper on learning from manipulation failures, now has a LeRobot conversion: 5,202 real ALOHA dual-arm episodes across 100 tasks, 4,545 of them failures, each failure with an annotated keyframe to avoid and one to correct. It is MIT-licensed; the depth maps were left out.
Figure has come out of stealth with Index, a robot training dataset it says is the most diverse built so far. In four months its creator network uploaded over 16 million videos from more than 100 countries, processing 30 minutes of video every second, with $15 million paid out to contributors. Figure crossed 264,000 app downloads and 44,000 weekly active users while still in stealth, and says it will spend over $1 billion on data and compute in the next twelve months.
Hydra-0 represents robot actions as pixel motion, giving one visual interface for learning action consequences across embodiments, tasks and video backbones. The best configuration cuts robot-motion error by 90.4% against an action-conditioned baseline, and an emergent inverse mode predicts robot motion from object flow taken from a human demonstration. On the RoboLab benchmark it reaches a Pearson correlation of 0.96 between replayed and reference success rates, and supports zero-shot composition.
The IG-10K dataset accompanies The Imitator Game benchmark, with over 20,000 paired human-robot episodes across more than 50 tasks and six domains. It spans four levels of imitation difficulty in real and simulated settings, in LeRobot format.
The first fully parallel framework for LiDAR bundle adjustment, built for maps larger than a GPU's memory: asynchronous data loading, bottom-up voxelisation to extract planar features, and a parallel increment solver. On large public datasets it runs up to ten times faster while matching the mapping accuracy of existing methods. The code is on GitHub.
Galaxy General Robotics' HumanTracker benchmark scores humanoid motion tracking against 2,500 reference clips retargeted to a 29-DoF Unitree G1-style body, in four families from daily motion to highly dynamic moves. It adds 6,000 human preference pairs comparing rollouts from four trackers — GMT, TWIST2, SONIC and Humanoid-GPT — which train a HumanScore reward model meant to catch foot skating and unstable support that per-frame error misses. It is Apache 2.0, with an ECCV 2026 paper.
Data vendor DecisionFacts has published an evaluation sample of its SO-101 arm teleoperation data: human-teleoperated manipulation episodes filmed from two cameras at 30 fps, with joint states and actions in LeRobot v3.0 format, and failure and recovery episodes labelled alongside the successes. The sample is licensed CC BY-NC 4.0 for research only; the full catalogue is licensed separately.
G0.5 puts reasoning and action in one autoregressive stream, so a single transformer decoder emits both under one objective rather than bolting a flow-matching action expert onto a frozen vision-language model. It combines a cross-embodiment action tokenizer, an interleaved chain-of-thought stream and a visual memory module carrying several seconds of history. Because reasoning and action share one set of weights, prompts steer action granularity, task horizon and out-of-distribution handling without further training.
Everyday Manipulation 3D publishes 1,513 raw RGB-D clips — 10.28 hours across four participants and ten manipulation tasks.
NVIDIA has released 1,215 human manipulation demonstrations from the TACO dataset, retargeted to bimanual Sharpa hands, with 121,500 robot episodes generated from them by its Video to Data pipeline. The 924.8 GB set, in Parquet and LeRobot v3 format under CC-BY-4.0, is meant to show what the pipeline outputs and to let users try its grounding step or downstream training.
CaryX AI has released 116 episodes of five people doing ten household tasks, recorded with an iPhone Pro's LiDAR, with metric depth, camera pose, MANO hand pose, per-hand object contact, masks and time-aligned language on every frame. It comes in LeRobot v3.0 format under a non-commercial CC-BY-NC-SA-4.0 licence, built for method development rather than broad generalisation claims.
Ego2Robot turns egocentric videos of people handling objects into robot training data by retargeting their actions, rendering robot arms into the footage and filtering for quality: 18,561 hours across 15 robot morphologies, which the authors call the largest such dataset. Pretraining on it alongside robot data improved out-of-distribution generalisation, including on real robots.
