Scaling Robot Memory for Complex Tasks

Summary & Methodology Analysis

RoboTTT addresses the limitations of current robot foundation models that struggle with long-horizon tasks due to their reliance on short-history input. By integrating Test-Time Training (TTT) layers into a Diffusion Transformer (a type of model that iteratively refines outputs to generate actions) action head, the system stores history as recurrent states using fast weights. These weights are updated via gradient descent during both training and inference. To ensure the model retains its original capabilities, the researchers implemented a learned tanh gating mechanism that manages the contribution of the TTT layers, allowing the system to scale its context effectively. To scale training context length, the authors used a recipe that combines sequence action forcing and Truncated Backpropagation Through Time (TBPTT), a technique that approximates gradient calculations over long sequences by cutting off updates after a set number of steps. Finally, they utilized DAgger Distillation, a method of transferring knowledge from expert policy rollouts while masking the imitation loss to only include human corrections. This ensures the model learns robustly from suboptimal data. The system was validated across various setups including GR00T N1.7, GDN, Pup Go Car, Circuit, and Gear Bot. While the method enables significant performance gains, the paper notes that scaling the training context length increases overall computational costs during training. Additionally, the model does not currently address every potential failure mode encountered during deployment.