Allen Institute launches “Two Blocks” simulation: 24-fold training speedup exposes spatial generalization fragility
Listen to this article
Read by Anchor
Video clips of a robot folding clothes or dispensing spare parts remain unable to explain what the machine has actually learned, or to diagnose the precise moment when the control arm fails against a unexpected obstacle. In experimental science, the study of complex systems starts with petri dishes and isolated environments that allow conditions to be changed rigorously and the outcome to be observed; today the Allen Institute for AI is transferring this laboratory logic to motion research through the open-source simulation framework “Two Blocks”, the first step of a larger project called “Foundational Worlds”.
The system consists of a robotic arm and wooden blocks on a work table, a setup that at first glance appears overly simple, yet forces models to combine abstract reasoning with precise motor skill to push, flip and arrange the blocks. The decisive advantage here is the automatic generation of training data without the need for remote manual operation, a process that previously consumed hundreds of human hours. The system relies on Nvidia’s open-source Isaac Lab to generate short motor skills via reinforcement learning and motion planning, using the simulator’s precise information about each block’s position and geometric angle before the task-and-motion planner assembles these skills into integrated sequences.
Virtual training no longer requires massive server arrays to generate physical skills.
The Two Blocks framework succeeds in generating and executing hundreds of plans in parallel on a single graphics processing unit, achieving a parallel speedup of twenty-four times real time. To link motion with sensory perception, the framework is geometrically divided into two levels: a high-level cognitive planning model such as “Molmo 2” that reads the scene visually and selects the block and required skill via a visual signal interface, and a low-level execution model such as “ManiFlow” that translates the commands into physical movements of the arm, then pauses briefly to request the next directive from the higher-level model.
The research surprise emerged when vision, language and motion models were subjected to strict spatial generalization tests.
Despite massive pre-training, our analytical results showed that models like “Pai 0.5”, “Molmo Act 2” and “ManiFlow” suffer from clear fragility when asked to transfer skills to spaces they have not previously encountered, or to extrapolate tower constructions of heights beyond the training range. In parallel experiments that linked advanced agents such as “Claude Opus 5.5” to the simulation tools via a model-context protocol, the generative agent solved thirty consecutive planning problems flawlessly, but then made sudden perceptual errors: while it analyzed the gripper’s rotation with geometric precision, it built the block tower upside-down on its side without noticing the visual flaw.
This distinction between planning ability and failure to perceive physical details directly impacts automation and logistics in Gulf ports and large distribution warehouses in Saudi Arabia, the United Arab Emirates and Egypt. Building self-operating sorting lines in the region no longer requires importing closed-source automation systems costing millions of dollars or equipping costly human-control platforms; using fast simulation tools allows engineering teams to train motion skills locally on modest computing hardware. However, the experiment warns against blind reliance on ready-made models, as the spatial gap demands generating diverse scenarios and finely tuning weights before risking the transfer of agents from screens to actual warehouse floors.