Credits & licenses

Baxter tactile signals

Data by Youssef Amine, Christian Gianoglio and Maurizio Valle, associated with Embedded real-time objects hardness classification for robotic grippers. Source: Zenodo V1.0, DOI 10.5281/zenodo.18246104, licensed under Creative Commons Attribution 4.0 (CC BY 4.0).

The preview redraws one recorded 80-sample, 16-channel tactile window. Each channel is offset by its first sample; sensors 2, 11 and 16 appear as curves, with all 16 channels in a signed heatmap. The replay reveals 10 stored samples per display second, then holds the last frame for one second; this is a display rate, not a claimed acquisition rate. No smoothing or invented sensor values are used. Poster and thumbnail are selected or cropped from the complete-data view.

The Baxter replay, poster and thumbnail on this site are shared under CC BY 4.0 with the above attribution and transformation notice. The data license is separate from any software license.

RH20T drawer demonstration

RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot, by Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu and Cewu Lu (Shanghai Jiao Tong University). Source: the official RH20T website and its task 29 drawer demonstration. The website and its source repository state Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0).

The original white task-caption strip was cropped and the camera image enlarged 2× using Lanczos resampling. All 260 consecutive source-video frames retain their original order and 30-frame/s presentation rate. No audio is included. The poster selects a frame; the thumbnail additionally crops and resizes that frame. This official overview demonstration is separate from the episode used in the reported evaluation and is not a model prediction.

The adapted RH20T replay, poster and thumbnail on this site are shared under CC BY-SA 4.0, retaining the above attribution and change notice. This statement covers these website-demo adaptations; it does not relicense the underlying RH20T dataset, whose terms depend on the source scene.

CoPhy scene demonstrations

CoPhy: Counterfactual Learning of Physical Dynamics, by Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori and Christian Wolf (ICLR 2020). Sources: the official project page and released repository. The upstream source code is distributed under GNU GPL version 3.

These are locally rendered demonstrations using the released scene and rendering materials, including floor textures. Collision and Balls are reconstructions with velocity replays; Blocktower is a separate surface-push simulation based on the released generator. The paired layout, annotations, camera framing, poster selection and thumbnail crops are presentation changes. The previews do not replace the evaluated dataset episodes. The code license is not presented as a blanket license for every upstream dataset or generated image.

Burgers released trajectory

Neural Operator Discovery Burgers Dataset, by Zituo Chen. Source: Zenodo, DOI 10.5281/zenodo.20372988, released under CC BY 4.0.

The preview redraws one released trajectory as a spatial curve and space–time field, using the RdBu_r display colormap and fixed signed color limits. All 101 stored states retain their original order and numerical values; they cover simulation time 0–5 seconds and are presented at 10 frames/s. The poster selects frame 50; the thumbnail renders the full space–time field with nearest-neighbor resizing. This is a data visualization, not original camera imagery or a learned prediction. The Burgers replay, poster and thumbnail are shared under CC BY 4.0 with this attribution and change notice.

FitzHugh–Nagumo solver illustration

The numerical model and display palette follow the DR2D generator in Neural Operator Discovery Code, by Zituo Chen. Source: Zenodo, DOI 10.5281/zenodo.20406332; this separate software record is released under CC BY 4.0.

The displayed trajectory was newly generated by a Python reimplementation of the released equations, periodic stencil and RK4 integration, with seed 42, k = 0.03 and β = 0.20. It is a visualization-only trajectory, not a released Burgers sample or an official FHN trajectory. The two numerical fields use the upstream generator’s summer colormap, fixed shared signed color limits and 101 stored states over simulation time 0–10 seconds, presented at 10 frames/s. Poster and thumbnail use the final stored state. The code-record license is cited for the generator source; no Burgers dataset license is inferred for this newly generated trajectory.

Overcooked rendering

Rendering uses work by the JaxMARL contributors (Alexander Rutherford et al.): Overcooked V2 visualizer and grid rendering routines. The upstream repository, originally published as FLAIROx/JaxMARL, is now hosted under bold-lab-ai and distributes this software under the Apache License 2.0.

Renderer imports were adapted to the local environment constants; raster drawing was retained and the GUI was unused. Two UtilityGreedy controllers generate the illustrative rollout. The complete 121-frame GIF was re-encoded with nearest-neighbor resizing and a demonstration label, preserving layout, frame order and 120 ms per frame; poster and thumbnail select or crop a frame. This is not a trained-policy result. The cited license describes the upstream renderer software and is not asserted as a universal license for generated footage.

Pendulum dynamics and render geometry

The dynamics use the released CaDM environment by Kimin Lee, Younggyo Seo, Seunghyun Lee, Honglak Lee and Jinwoo Shin, Context-aware Dynamics Model for Generalization in Model-Based Reinforcement Learning (ICML 2020): official source repository. The display reproduces the geometry of OpenAI Gym 0.16 Pendulum; Gym is copyright 2016 OpenAI and is distributed under the MIT License.

This is a new fixed-input torque demonstration using unchanged CaDM reset/step and normalization methods. The original Gym geometry and colors were redrawn with Pillow, with torque and response overlays; these are not native Gym-rendered pixels. The 161 frames are shown at 20 frames/s, with a selected poster and cropped thumbnail. Gym’s license is not extended to CaDM, whose linked repository does not supply a top-level license, and no separate CaDM license is asserted here.

Project-generated previews

D-Clean redraws stored numerical states and actual force samples at half physical speed, with force and drag overlays. PokeWorld replays the recorded method-figure histories at quarter speed with normalized-action arrows and trajectory overlays. SpringWorld and Swimmer use new illustrative input sequences in the project’s simulation models, with force or torque annotations, selected posters and cropped thumbnails. These previews do not add evaluation results.

SpringWorld and Swimmer use MuJoCo by the MuJoCo contributors, distributed as software under the Apache License 2.0. These are project-generated models and views; citing the simulation software does not assign its license to every generated image.