NTT's NTTワールドモデル boosts robot task success to 92.3%

NTT's announcement addresses a known weakness in world models: systems that learn future changes empirically from large amounts of video data can generate predictions that violate physical laws in situations they have not learned. That is especially problematic at sites handling airport baggage or parcel deliveries, where shapes, weights and placements vary and collecting training data for every situation is difficult. NTT's approach pairs a video-generation model with a proprietary 'physics control adapter' that estimates object position, orientation, velocity and mass and predicts state changes based on equations of motion. By integrating the experience-based generator with the physics-based adapter, NTT says it suppresses unnatural predictions such as objects disappearing or deforming. The technology also works with existing robot control software, predicting post-execution situations in video for multiple action candidates from VLA models and motion planners and scoring each action's success, so a robot can compare options before moving. Beyond the shelf-stacking result, NTT reported a LPIPS score of 7.81 on the OBJ3D benchmark, above OCK's 7.92 and NVIDIA Cosmos Predict 2.5's 21.90. The first intended application is airport baggage handling, with verification to proceed with Mujin and expansion targeted at logistics, warehousing, manufacturing, construction and retail.