# World Action Model

World Action Models (WAMs) predict future visual states and actions together, but generating a visual future at every replanning step is expensive and sometimes unnecessary. Bind-WAM uses whether the roles named in a language instruction are stably bound to objects in the current view as an interpretable signal: an offline pipeline labels reference mismatch, invisibility and identity drift from complete demonstrations, and a lightweight router of about 6.65M parameters switches a shared WAM between direct action prediction and joint future-action denoising. On LIBERO, Bind-WAM reaches 98.60% mean success while invoking the joint path at only 40.58% of replanning states, cutting average completion time from 15.73 s for Fast-WAM-Joint to 11.71 s.

World Action Models (WAMs) predict future visual states and actions together, but generating a visual future at every replanning step is expensive and sometimes unnecessary. Bind-WAM uses whether the roles named in a language instruction are stably bound to objects in the current view as an interpretable signal: an offline pipeline labels reference mismatch, invisibility and identity drift from complete demonstrations, and a lightweight router of about 6.65M parameters switches a shared WAM between direct action prediction and joint future-action denoising. On LIBERO, Bind-WAM reaches 98.60% mean success while invoking the joint path at only 40.58% of replanning states, cutting average completion time from 15.73 s for Fast-WAM-Joint to 11.71 s.

Shanghai Jiao Tong University and collaborators assembled a roughly 47,000-hour multisource pool and extended a pretrained video model's language–video priors into language–video–action learning, yielding the world action model SWAM. On the 50 manipulation tasks of RoboTwin 2.0, SWAM reaches 61.56% success under the randomized Hard setting, 12.86 percentage points above the strongest baseline OpenWAM, and 88.10% under the clean Easy setting, slightly below OpenWAM's 89.30%. On a real bimanual platform, its three-task average Hard success rate is 40.00% against OpenWAM's 23.33%.

Shanghai Jiao Tong University and collaborators assembled a roughly 47,000-hour multisource pool and extended a pretrained video model's language–video priors into language–video–action learning, yielding the world action model SWAM. On the 50 manipulation tasks of RoboTwin 2.0, SWAM reaches 61.56% success under the randomized Hard setting, 12.86 percentage points above the strongest baseline OpenWAM, and 88.10% under the clean Easy setting, slightly below OpenWAM's 89.30%. On a real bimanual platform, its three-task average Hard success rate is 40.00% against OpenWAM's 23.33%.

At CoRL, NVIDIA released an open robotics stack including GR00T 2, a world action model built on Cosmos, the Isaac Lab 3.0 simulation framework and the Isaac Lab-Arena evaluation framework. Tighter integration between Isaac Lab 3.0 and the open-source Newton physics engine lets it model deformable objects such as cables and cloth and multi-joint robotic hands; on a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reached up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. GR00T 2 averaged 71% success across 50 tasks with one demonstration per task, rising to 95% with five.

At CoRL, NVIDIA released an open robotics stack including GR00T 2, a world action model built on Cosmos, the Isaac Lab 3.0 simulation framework and the Isaac Lab-Arena evaluation framework. Tighter integration between Isaac Lab 3.0 and the open-source Newton physics engine lets it model deformable objects such as cables and cloth and multi-joint robotic hands; on a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reached up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. GR00T 2 averaged 71% success across 50 tasks with one demonstration per task, rising to 95% with five.

Sber Robotics Center released GreenWAM, a robot foundation model rebuilt from Cosmos 3 Edge to serve as both a world model and a policy. A shared 79-slot anatomical action space maps the native joints of 20 robot families onto common leg, torso, arm and finger coordinates, and one model is trained jointly in seven visual and action modes. On WorldArena 1.0 it has the highest trajectory, depth and 3D accuracy among the compared world-action models and video generators; on the real Green humanoid it completes 95.6% of in-distribution instructions, against 88.9% for a policy adapted from Cosmos 3 Edge and 84.4% for π0.5.

Sber Robotics Center released GreenWAM, a robot foundation model rebuilt from Cosmos 3 Edge to serve as both a world model and a policy. A shared 79-slot anatomical action space maps the native joints of 20 robot families onto common leg, torso, arm and finger coordinates, and one model is trained jointly in seven visual and action modes. On WorldArena 1.0 it has the highest trajectory, depth and 3D accuracy among the compared world-action models and video generators; on the real Green humanoid it completes 95.6% of in-distribution instructions, against 88.9% for a policy adapted from Cosmos 3 Edge and 84.4% for π0.5.

GUI agents no longer reconstruct the next screen pixel by pixel; a shared latent representation now scores candidate actions and predicts action-conditioned futures at once, raising success rates by up to 31.9% offline and 40.6% online relative to pixel-space world models while cutting world-model inference latency by up to 98.9%.

GUI agents no longer reconstruct the next screen pixel by pixel; a shared latent representation now scores candidate actions and predicts action-conditioned futures at once, raising success rates by up to 31.9% offline and 40.6% online relative to pixel-space world models while cutting world-model inference latency by up to 98.9%.

A world action model can regenerate complete rollouts of previously learned tasks from their instructions alone, and after truncation and physical-consistency filtering, replaying them cuts the forgetting of sequential fine-tuning by over 90% on four LIBERO suites and five real-world tasks without keeping any data from previous tasks.

A world action model can regenerate complete rollouts of previously learned tasks from their instructions alone, and after truncation and physical-consistency filtering, replaying them cuts the forgetting of sequential fine-tuning by over 90% on four LIBERO suites and five real-world tasks without keeping any data from previous tasks.

SmolWAM compresses the 2B-parameter Cosmos Policy video world-action model to 0.88B, succeeding in 95.5% of 400 LIBERO episodes at 57 ms, 2.3 TFLOP and 2.4 GiB per 16-action chunk, 11 times faster than its teacher while still forecasting scene motion.

SmolWAM compresses the 2B-parameter Cosmos Policy video world-action model to 0.88B, succeeding in 95.5% of 400 LIBERO episodes at 57 ms, 2.3 TFLOP and 2.4 GiB per 16-action chunk, 11 times faster than its teacher while still forecasting scene motion.