datasets.bot — Curated catalog of robotics and embodied-AI training datasets

TL;DR

datasets.bot is a browsable catalog of robotics and embodied-AI training datasets, tagging each entry with provider, year, license, modality coverage (rgb, depth, ee_pose, language, proprioception, point_cloud, force_torque, tactile, ego), popularity signal (download count, GitHub stars, or citation count), and a star/featured marker. The landing page exposes a mix of recent 2026 releases — LeRobot Shirt-Folding, FTP-1, HIW-500 (BitRobot), AgiBot World 2026, ABC-130k, NVIDIA GR00T X-Embodiment Sim, NVIDIA DreamDojo GR-1 Post-Training, an Ai2 dataset — alongside license-and-modality filters. Useful as an external “what robot/embodied dataset exists, with what license, and how popular is it” dashboard when scoping a VLA or world-model project.

Key claims

• Indexes embodied/robotics datasets across providers (LeRobot/Hugging Face, NVIDIA, BitRobot, AgiBot/Zhiyuan Robotics, FTP-1 authors, multi-institution academic releases, Ai2), with each entry annotated by year, license, and modality tag set [landing page listings].
• Surfaces popularity proxies per dataset — download counts (e.g. 17k, 16k, 7.6k), citation counts (e.g. 345, 55), and GitHub star counts (e.g. 149★) — letting a reader sort by adoption signal rather than recency alone [landing page entries].
• Tags every dataset with explicit modality coverage from a shared vocabulary (rgb, depth, ee_pose, language, proprioception, point_cloud, force_torque, tactile, ego), enabling filtering by sensor mix when matching a dataset to a target embodiment or training objective [landing page entries].
• Includes license terms inline (Apache-2.0, MIT, CC-BY-NC-4.0, custom, unknown) so non-commercial-only datasets are flagged at browse time rather than at download time [landing page entries].

Method

The site is a curated index, not a host: each entry links out to the source provider (Hugging Face, GitHub, etc.) and reports metadata fetched from those sources. Authorship of the catalog itself is undisclosed on the landing page; the utm_source=substack parameter on the link suggests it is promoted via a Substack newsletter. Content type is text/html but the page returned no extractable body to Bud’s fetcher — likely a client-rendered SPA — so the entries summarized above were recovered from third-party indexing. Treat the catalog as a starting-point browse aid, not as ground truth: provider-side download counts and citation counts are noisy popularity proxies, and the ”★” featured marker is unexplained.

Results

Not a research artifact — no benchmark numbers to report. The site itself is the artifact. As a snapshot of what the catalog highlights as of 2026-06-30: the visible landing entries skew heavily toward 2026 robot manipulation + humanoid data (LeRobot Shirt-Folding, AgiBot World 2026, HIW-500 humanoids-in-the-wild, NVIDIA GR00T X-Embodiment Sim, ABC-130k, FTP-1), with ego, rgb, proprioception, ee_pose, and language as the most common modality tags. tactile and force_torque appear on at least one entry (FTP-1) but are notably rarer than visual + proprioceptive modalities — consistent with the wiki’s standing observation that tactile/force data is a known gap in the embodied-AI dataset landscape.

Why it’s interesting

datasets.bot is the robotics/embodied-AI counterpart to Sophon — AI Research Catalog of Evals, Tools, and Labs (which catalogs evals/tools/labs) and awesome-diffusion-papers.vercel.app — time-sorted, searchable browser for awesome-diffusion-categorized (which catalogs diffusion papers): a third-party browseable index rather than a research artifact. It complements the team’s standing interest in the human-vs-robot-data substrate debate filed across Human-to-Robot Retargeting, HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining, EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World, and EgoInfinity: A Web-Scale 4D Hand-Object Interaction Data Engine for Any-View Robot Retargeting and Video-to-Action Robot Learning — when “should we train on dataset X or Y” comes up, this is the dashboard for “what X and Y even exist, under what license, with what modality coverage.” Also useful as a sanity check on the licence-and-modality structure of NVIDIA’s DreamDojo GR-1 Post-Training release (CC-BY-NC-4.0, ego/rgb/language/proprioception/ee_pose) and other recent drops that the wiki has filed individually but not aggregated.

See also

• Sophon — AI Research Catalog of Evals, Tools, and Labs — same third-party-catalog pattern for evals/tools/labs rather than datasets.
• awesome-diffusion-papers.vercel.app — time-sorted, searchable browser for awesome-diffusion-categorized — same pattern for diffusion papers.
• Survey notes on World Action Models / VLA for robotics (index) — narrative survey of the WAM/VLA corpus; datasets.bot is the dataset-side complement.
• Human-to-Robot Retargeting — the bridging-data debate this catalog can be used to scope.
• Synthetic Training Data — datasets.bot indexes synthetic releases (NVIDIA GR00T X-Embodiment Sim, DreamDojo) alongside real-world captures, useful for comparing the two regimes at browse time.
• VLA Models — the dominant training target for the catalog’s entries.