Papers
34 publications that produced the datasets, benchmarks and policies in the catalog. Every paper links back to the robots, tasks, datasets and models it connects.
Chi, Feng, Du et al. · Columbia / TRI / MIT · RSS 2023 · arXiv:2303.04137
Represents the visuomotor policy as a conditional denoising diffusion process over action sequences, improving average success by 46.9% over prior behavior cloning baselines.
SourceBrohan et al. · Google Robotics · RSS 2023 · arXiv:2212.06817
A 35M-parameter transformer policy trained on 130k real teleoperated episodes across 700+ kitchen tasks, showing that model capacity plus data diversity drives generalization on real hardware.
SourceBrohan et al. · Google DeepMind · CoRL 2023 · arXiv:2307.15818
Co-fine-tunes internet-scale vision-language models on robot trajectories, emitting actions as text tokens and inheriting semantic reasoning about unseen objects and instructions.
SourceOpen X-Embodiment Collaboration · 53 institutions · ICRA 2024 · arXiv:2310.08864
Pools 1M+ real robot trajectories from 22 embodiments into a single RLDS collection and shows positive transfer when training a shared policy across all of them.
SourceZhao, Kumar, Levine, Finn · Stanford · RSS 2023 · arXiv:2304.13705
Introduces ALOHA teleoperation and Action Chunking with Transformers, learning contact-rich bimanual skills from 50 demonstrations.
SourceMandlekar, Xu, Wong et al. · Stanford / NVIDIA · CoRL 2021 · arXiv:2108.03298
A large study of offline learning from proficient and mixed-quality human demonstrations, releasing the robomimic datasets and reproducible baselines.
SourceKim, Pertsch, Karamcheti et al. · Stanford / Berkeley / TRI / Google DeepMind · CoRL 2024 · arXiv:2406.09246
A 7B open VLA trained on 970k Open X-Embodiment episodes that outperforms RT-2-X with 7x fewer parameters and supports LoRA fine-tuning on consumer GPUs.
SourceOcto Model Team · Berkeley RAIL / Stanford · RSS 2024 · arXiv:2405.12213
A transformer diffusion policy pretrained on 800k Open X-Embodiment trajectories, designed for cheap fine-tuning to new sensors, action spaces and embodiments.
SourceJang, Irpan, Khansari et al. · Google / Everyday Robots · CoRL 2021 · arXiv:2202.02005
25k demonstrations plus 11k human-video interventions over 100 tasks, conditioned on language or video embeddings for zero-shot transfer to new tasks.
SourceFu, Zhao, Finn · Stanford · CoRL 2024 · arXiv:2401.02117
Extends ALOHA to a wheeled base with whole-body teleoperation and shows co-training with static ALOHA data lifts mobile manipulation success up to 90%.
SourceWalke, Black, Zhao et al. · Berkeley RAIL · CoRL 2023 · arXiv:2308.12952
60k trajectories over 24 environments on a WidowX arm, with language annotations, designed to study broad generalization for scalable imitation learning.
SourceMees, Hermann, Rosete-Beas, Burgard · University of Freiburg · RA-L 2022 · arXiv:2112.03227
A simulation benchmark of 34 tasks across four environments where policies must chain instructions from unstructured play data and raw pixels.
SourceBlack, Brown, Driess et al. · Physical Intelligence · preprint 2024 · arXiv:2410.24164
Adds a flow-matching action expert to a VLM backbone for 50Hz dexterous control, trained on a large cross-embodiment mixture plus in-house dexterous data.
SourceLiu, Zhu et al. · UT Austin · NeurIPS 2023 · arXiv:2306.03310
130 language-conditioned simulation tasks in four suites that separate spatial, object, goal and long-horizon knowledge transfer for lifelong robot learning.
SourceDasari, Ebert, Tian et al. · Berkeley / Stanford / CMU / Penn · CoRL 2019 · arXiv:1910.11215
15M video frames of self-supervised interaction across seven robot platforms, used to pretrain visual dynamics models that transfer to unseen hardware.
SourceKhazatsky, Pertsch et al. · Multi-institution consortium · RSS 2024 · arXiv:2403.12945
76k teleoperated trajectories collected in 564 real scenes by 50 operators across 13 institutions, all on the same Franka + handheld-controller rig.
SourceGu, Xiang, Li et al. · UC San Diego · ICLR 2023 · arXiv:2302.04659
20 manipulation task families with 2000+ object variations, 4M demonstration frames and a GPU-accelerated simulator running thousands of FPS.
SourceMandlekar, Nasiriany et al. · NVIDIA · CoRL 2023 · arXiv:2310.17596
Generates 50k+ synthetic demonstrations from ~200 human demos by transforming object-centric subtask segments to new scenes and objects.
SourceLynch, Wahid, Tompson et al. · Google Robotics · RA-L 2023 · arXiv:2210.06407
Releases Language-Table, the largest real-world language-conditioned manipulation dataset at the time, and a policy steerable by live natural-language guidance.
SourceHoeller, Rudin, Sako, Hutter · ETH Zurich RSL · Science Robotics 2024 · arXiv:2306.14874
A hierarchy of RL skills plus a perceptive navigation policy that lets ANYmal jump gaps, climb obstacles and traverse parkour terrain from onboard depth.
SourceHansen, Su, Wang · UC San Diego · ICLR 2024 · arXiv:2310.16828
A model-based RL algorithm that plans in a learned implicit latent space, scaling to a single 317M-parameter agent solving 80 continuous-control tasks.
SourceBharadhwaj, Vakil, Sharma et al. · CMU / Meta AI · ICRA 2024 · arXiv:2309.01918
Releases RoboSet, 100k trajectories of 12 kitchen skills, and shows semantic augmentation plus action chunking generalizes to 38 tasks from modest data.
SourceFang, Fang, Wang et al. · Shanghai Jiao Tong University · ICRA 2024 · arXiv:2307.00595
110k contact-rich robot episodes with force-torque, multi-view RGB-D and paired human demonstrations across 147 tasks and seven robot configurations.
SourceKe, Gkanatsios, Fragkiadaki · Carnegie Mellon University · CoRL 2024 · arXiv:2402.10885
Marries 3D scene tokens with diffusion policies for language-conditioned manipulation, setting new state of the art on RLBench and CALVIN.
SourceMees, Borja-Diaz, Burgard · University of Freiburg · ICRA 2023 · arXiv:2210.01911
Learns language-conditioned affordances from unstructured play data and free-form instructions, the setting behind the TACO-Play dataset.
SourceWang, Xu, Zhu et al. · MIT CSAIL · NeurIPS 2024 · arXiv:2409.20537
Aligns proprioception and vision from 52 heterogeneous datasets into a shared latent token space and shows scaling behaviour for a transferable policy trunk.
SourceLiu, Chen, Zhu et al. · Tsinghua University · ICLR 2025 · arXiv:2410.07864
A 1B-parameter diffusion transformer pretrained on 1M+ multi-robot episodes with a unified action space, then fine-tuned on 6k self-collected bimanual episodes.
SourceALOHA 2 Team · Google DeepMind · tech report 2024 · arXiv:2405.02292
A hardware revision of ALOHA with improved grippers, gravity compensation and durability, plus a MuJoCo simulation model for large-scale data collection.
SourcePertsch, Stachowicz, Ichter et al. · Physical Intelligence · preprint 2025 · arXiv:2501.09747
A DCT-based action tokenizer that compresses high-frequency action chunks, letting autoregressive VLAs train 5x faster while matching diffusion-based control quality.
SourceNVIDIA · NVIDIA GEAR · preprint 2025 · arXiv:2503.14734
A dual-system VLA for humanoids that mixes real teleoperation, synthetic neural trajectories and human video into one pretraining recipe.
SourceCadene, Alibert, Soare et al. · Hugging Face · open-source library 2024
An open library of datasets, pretrained policies and low-cost hardware designs (SO-100/SO-101) that standardizes robot learning workflows in PyTorch.
SourceAgiBot-World Contributors · AgiBot · preprint 2025 · arXiv:2503.06669
1M+ long-horizon episodes on a standardized dual-arm humanoid fleet with dexterous hands, plus the GO-1 policy baseline.
SourceKim et al. · Stanford · preprint 2025 · arXiv:2502.19645
Parallel decoding, action chunking and continuous action heads raise fine-tuned VLA throughput 25x and lift bimanual ALOHA success to 97%.
SourceShukor, Aubakirova, Capuano et al. · Hugging Face · preprint 2025 · arXiv:2506.01844
A 450M VLA trained only on community-contributed LeRobot datasets that runs on a single consumer GPU and supports asynchronous inference on low-cost arms.
Source