Published policies — vision-language-action models, diffusion policies and world-model agents — with the data they were trained on, the compute it took and the benchmarks their authors report.
OpenVLA 7B
7BFully open 7B vision-language-action model trained on 970k real robot episodes. Beats RT-2-X by 16.5% absolute on 29 tasks with 7x fewer parameters, and fine-tunes on one A100 with LoRA.
Prismatic VLM (Llama-2 7B + DINOv2/SigLIP) · Franka / WidowX / mobile manipulators · 970K train episodes · 96K downloads
Diffusion Policy (UNet)
110MFormulates visuomotor control as conditional denoising diffusion over action sequences, with receding-horizon replanning. Averaged a 46.9% relative improvement over prior state of the art across 15 tasks.
Conditional DDPM visuomotor policy · Franka / UR5 / mobile bimanual · 5.5K train episodes · 93K downloads
π0 (pi-zero)
3.3BFlow-matching vision-language-action model producing 50 Hz action chunks. Trained on 10k+ hours across 7 robot platforms and 68 tasks; folds laundry, bags groceries and assembles boxes end to end.
Flow-matching VLA (PaliGemma + action expert) · Bimanual arms, mobile manipulators, ALOHA · 900K train episodes · 88K downloads
ACT (Action Chunking Transformer)
80MThe policy behind ALOHA: a CVAE transformer predicting chunks of ~100 future actions, learning millimetre-precision bimanual skills from only 50 demonstrations per task.
CVAE transformer with action chunking · ALOHA low-cost bimanual · 1.2K train episodes · 79K downloads
Octo-Base
93MOpen generalist policy trained on 800k Open X-Embodiment trajectories with a diffusion action head; supports new sensors and action spaces by adding lightweight input/output heads.
Diffusion-head transformer policy · Multi-embodiment (WidowX, Franka, mobile) · 800K train episodes · 71K downloads
SmolVLA-450M
450MSmall, community-data VLA that runs on a consumer GPU or CPU, trained entirely on ~23k public LeRobot community episodes yet matching much larger VLAs on real low-cost arms.
Compact VLA (SmolVLM-2 + flow-matching expert) · SO-100 / SO-101 / Koch low-cost arms · 23K train episodes · 65K downloads
GR00T N1-2B
2.2BOpen humanoid foundation model with a System-2 VLM reasoner and System-1 diffusion transformer for 120 Hz motor commands, trained on real teleoperation, simulation rollouts and neural-video-generated trajectories.
Dual-system VLA (Eagle-2 VLM + diffusion transformer) · Humanoid (Fourier GR-1, Unitree) + bimanual arms · 780K train episodes · 59K downloads
π0-FAST
3BReplaces per-dimension binning with a DCT-based action tokenizer, cutting token count 5x and making autoregressive VLAs viable at 50 Hz control rates.
Autoregressive VLA with FAST action tokenizer · Bimanual + mobile manipulators · 850K train episodes · 43K downloads
RT-2-X 55B
55BCo-fine-tunes a web-scale vision-language model on robot trajectories, emitting actions as discretized text tokens. RT-2-X adds the cross-embodiment Open X-Embodiment mixture for emergent generalization.
Vision-language-action transformer (PaLI-X backbone) · Mobile manipulator · 1.1M train episodes · 41K downloads
OpenVLA-OFT
7BOptimized fine-tuning recipe for OpenVLA: parallel action-chunk decoding, continuous L1 regression head and FiLM language conditioning give 26x faster inference and near-saturated LIBERO scores.
OpenVLA + parallel decoding & continuous actions · ALOHA 2 / Franka · 62K train episodes · 34K downloads
RDT-1B
1.2BLargest open diffusion foundation model for bimanual manipulation, pre-trained on 1M+ multi-robot episodes with a unified physically-interpretable action space, then fine-tuned on 6k self-collected bimanual episodes.
Diffusion transformer with unified action space · ALOHA / dual-arm · 1.0M train episodes · 34K downloads
RT-1
35MThe original Robotics Transformer: 35M parameters, 3 Hz closed-loop control, trained on 130k teleoperated episodes across 700+ kitchen tasks.
EfficientNet-B3 + TokenLearner + Transformer · Everyday Robot mobile manipulator · 130K train episodes · 29K downloads
TD-MPC2
317MSingle 317M agent trained on 104 continuous-control tasks across 4 domains, planning in a learned implicit world model. Scales predictably with model and data size.
Model-based RL with latent-space planning · Quadruped, humanoid, manipulator (sim) · 27K downloads
3D Diffuser Actor
180MMarries 3D scene representations with diffusion policies: lifts multi-view features into 3D tokens and denoises end-effector trajectories, setting state of the art on RLBench and CALVIN.
3D scene-token diffusion policy · Franka Panda (RLBench / CALVIN) · 3.2K train episodes · 21K downloads
HPT-XLarge
1.1BLearns a shared policy trunk across 52 heterogeneous datasets by tokenizing each robot proprioception and vision stream into a common latent, then attaching lightweight task heads.
Heterogeneous pre-trained transformer trunk · 52 embodiments (arms, mobile, sim) · 200K train episodes · 19K downloads