Disentangled Robot Learning via Separate Forward and Inverse Dynamics Pretraining

A research team has introduced DeFI (Disentangled Forward and Inverse dynamics pretraining), a novel framework designed to overcome a core limitation in current robot AI. Most advanced robots use Vision-Language-Action (VLA) models, but these entangle the tasks of predicting future video frames (forward dynamics) and deciding on actions (inverse dynamics). This entanglement creates a misalignment between 2D image forecasting and 3D action prediction and, crucially, prevents models from learning from the vast amounts of action-free video data available on the web.

DeFI solves this by splitting the problem into two specialized, pre-trainable components. First, a General Forward Dynamics Model (GFDM) is trained on massive datasets of diverse human and robot videos purely to predict future visual states. Second, a General Inverse Dynamics Model (GIDM) uses self-supervised learning to infer the latent actions that caused transitions between video frames. These two expert models are then integrated and fine-tuned together for specific robotic tasks. This 'pretrain separately, cooperate later' approach allows each model to learn from its optimal data source before combining their strengths.

The results are state-of-the-art. On the challenging CALVIN benchmark, DeFI achieved an average task length of 4.51. It scored a 51.2% success rate on the SimplerEnv-Fractal benchmark and, most impressively, an 81.3% success rate in real-world robotic deployments. These figures represent a significant leap over previous entangled training methods. The work, accepted at ICLR 2026, demonstrates a more data-efficient and performant pathway toward generalist robots capable of complex, long-horizon reasoning by fundamentally rethinking how they learn from visual data.