Entries newly merged into the WAM atlas by the daily agent pipeline tag:wangevan6.github.io,2026:awesome-world-action-models 2026-07-21T00:00:00Z awesome-world-action-models tag:wangevan6.github.io,2026:awesome-world-action-models/robottt 2026-07-21T00:00:00Z

[vla] Integrates test-time training into robot foundation policies, compressing long visuomotor histories into gradient-updated fast weights so a policy can imitate from a single human video and improve on the fly.

tag:wangevan6.github.io,2026:awesome-world-action-models/rynnbrain-1-1 2026-07-21T00:00:00Z

[vla] Open family of embodied foundation models for perception, spatial reasoning, localization, and planning, with a VLA variant that maps a shared cross-embodiment action space onto several real robots.

tag:wangevan6.github.io,2026:awesome-world-action-models/xiaomi-robotics-1 2026-07-21T00:00:00Z

[vla] Pretrains a vision-language-action model on real-world manipulation trajectories collected with handheld UMI devices and auto-labeled with scene-transition language, then post-trains it to follow instructions on robots.

tag:wangevan6.github.io,2026:awesome-world-action-models/cosmos-3 2026-07-21T00:00:00Z

[unified-wam] An open Mixture-of-Transformers omnimodel over language, image, video, audio, and action that switches between multimodal reasoning, video generation, and joint action prediction with world simulation.

tag:wangevan6.github.io,2026:awesome-world-action-models/pi0-7 2026-07-21T00:00:00Z

[vla] A 5B steerable generalist VLA conditioned on multimodal context — language, performance metadata, visual subgoals — that matches specialist models out of the box and transfers zero-shot across embodiments.

tag:wangevan6.github.io,2026:awesome-world-action-models/dreamzero 2026-07-21T00:00:00Z

[unified-wam] Builds a World Action Model on a pretrained video-diffusion backbone that jointly predicts future states and actions, acting as a zero-shot policy on new tasks without task-specific training.

tag:wangevan6.github.io,2026:awesome-world-action-models/ctrl-world 2026-07-21T00:00:00Z

[unified-wam] A multi-view, frame-level action-conditioned world model with pose-conditioned memory that lets generalist VLA policies roll out, be evaluated, and be improved entirely in imagination.

tag:wangevan6.github.io,2026:awesome-world-action-models/wall-oss 2026-07-21T00:00:00Z

[vla] An open embodied foundation model that tightly couples a VLM with action synthesis via a Unified Cross-Level chain-of-thought spanning reasoning, subgoal decomposition, and control.

tag:wangevan6.github.io,2026:awesome-world-action-models/genie-3 2026-07-21T00:00:00Z

[video-world-models] A general-purpose world model that generates navigable, promptable 720p worlds in real time at 24 fps, staying consistent for minutes and supporting on-the-fly 'promptable world events'.

tag:wangevan6.github.io,2026:awesome-world-action-models/genie-envisioner 2026-07-21T00:00:00Z

[unified-wam] A single video-generative world foundation platform (GE-Base/GE-Act/GE-Sim) that unifies policy learning, action execution, and neural simulation for real robots, trained on 1M+ AgiBot trajectories.

tag:wangevan6.github.io,2026:awesome-world-action-models/dreamvla 2026-07-21T00:00:00Z

[unified-wam] Has the policy 'dream' compact world knowledge — dynamic regions, depth, and semantics — instead of full frames, then decodes actions from the dream with a diffusion head.

tag:wangevan6.github.io,2026:awesome-world-action-models/hunyuan-gamecraft 2026-07-21T00:00:00Z

[video-world-models] Generates high-dynamic interactive game videos by unifying keyboard and mouse inputs into a shared camera-representation space, with history-conditioned autoregressive extension.

tag:wangevan6.github.io,2026:awesome-world-action-models/matrix-game 2026-07-21T00:00:00Z

[video-world-models] A 17B open-source image-to-world model with fine-grained keyboard/mouse control, trained on 2,700+ hours of Minecraft data, outperforming Oasis and MineWorld on controllability and physics.

tag:wangevan6.github.io,2026:awesome-world-action-models/roboarena 2026-07-21T00:00:00Z

[evaluation] A Chatbot-Arena-style crowd-sourced network of real-robot evaluators running double-blind A/B comparisons of generalist policies on freely chosen tasks, aggregated into policy rankings.

tag:wangevan6.github.io,2026:awesome-world-action-models/smolvla 2026-07-21T00:00:00Z

[vla] A 450M-parameter VLA trained on community LeRobot datasets that matches models 10x its size and runs on consumer GPUs or even CPUs, with async inference.

tag:wangevan6.github.io,2026:awesome-world-action-models/v-jepa-2 2026-07-21T00:00:00Z

[unified-wam] A self-supervised video model trained on 1M+ hours of internet video whose action-conditioned variant V-JEPA 2-AC, post-trained on only 62 hours of robot video, plans pick-and-place zero-shot on new Franka arms.

tag:wangevan6.github.io,2026:awesome-world-action-models/worldvla 2026-07-21T00:00:00Z

[unified-wam] Fuses a VLA and a world model into one autoregressive transformer where action generation and future-image prediction mutually improve each other, with an attention mask fixing action-chunk error accumulation.

tag:wangevan6.github.io,2026:awesome-world-action-models/dreamgen 2026-07-21T00:00:00Z

[unified-wam] Finetunes video world models to dream photorealistic robot trajectories, labels them with latent or inverse-dynamics actions, and trains policies on the resulting 'neural trajectories'.

tag:wangevan6.github.io,2026:awesome-world-action-models/ewmbench 2026-07-21T00:00:00Z

[evaluation] A benchmark and toolkit scoring embodied (robot) world models on visual scene consistency, motion/trajectory correctness, and semantic alignment with language instructions.

tag:wangevan6.github.io,2026:awesome-world-action-models/pi0-5 2026-07-21T00:00:00Z

[vla] Extends π0 with heterogeneous co-training and high-level subtask inference so mobile manipulators can clean kitchens and bedrooms in entirely unseen homes.