A 4D geometry supervisor (3D space plus time) trains a robot foundation model during training, then steps aside before the robot runs, with paper reported benchmark gains and no added compute.
A new training recipe for robot foundation models lets a robot learn to reason about how objects move through three-dimensional space over time, then drops the geometry teacher before the robot ever runs. The result, called MECo-WAM and reported in a recent arXiv preprint, posts near-state-of-the-art scores on two standard manipulation benchmarks while keeping the deployed model the same size and compute footprint as the baseline.
The work sits inside two fast-moving ideas in robot learning: video-action co-training, where one model learns to predict what a robot's camera will see and what its joints should do at the same time, and 4D geometric priors, the explicit use of three-dimensional shape plus time as a planning signal. MECo-WAM (Multi-Expert Co-Training World Action Model) is a video-action model for robotic manipulation. "4D" here means 3D geometry plus time, not a fourth spatial axis.
Video-action co-training has been tried before, with mixed results. When a single model is asked to predict pixel-like video frames and motor commands simultaneously, the visual side tends to dominate. The model ends up optimizing for what an object looks like in still frames rather than how it shifts, rotates, and collides over time. That bias shows up at the gripper: a model that knows what a cup looks like but not how it tips under torque will knock the cup over.
MECo-WAM adds a third "expert" alongside the existing video and action experts: a lightweight 4D module supervised by relational geometry targets from a frozen VGGT encoder, a 3D-vision model the paper borrows but does not retrain. The frozen teacher runs at training time and nowhere else; the deployed graph stays the same size and compute as the baseline.
Two design choices carry the result. First, asymmetric expert visibility: the auxiliary 4D expert can see and learn from the video and action experts, but the deployed video-action pathway cannot see the 4D expert's outputs directly. That arrangement prevents non-causal geometry from leaking into the deployed model's decisions. Second, a decayed 4D read-mask attention: early in training, the deployed pathway attends to geometric guidance from the 4D expert; as training progresses, that attention is masked out, so the deployed pathway ends up holding its spatial reasoning internally rather than borrowing from a runtime module. A third component, action-aware temporal geometric distillation, adds a regularizer that keeps the internal geometry consistent with action sequences the robot could actually execute.
On the manipulation benchmarks, the gains are concrete. MECo-WAM reports 98.2% on LIBERO and 92.6% on RoboTwin 2.0, both standard test suites, while the deployed inference graph matches the baseline in size and compute. The "no inference cost increase" claim is the load-bearing piece: it depends on the auxiliary 4D modules being cleanly removable, which the paper asserts and the field will want to verify independently.
Three caveats sit underneath these numbers. First, LIBERO and RoboTwin 2.0 are research benchmarks with curated scenes, not deployment evidence on warehouse floors or kitchen counters. Second, VGGT is referenced as a frozen encoder, but its origin and training scope are not separately corroborated outside this paper. And video-action co-training is a young paradigm with known failure modes around appearance-style latents; MECo-WAM's asymmetric expert visibility is itself an engineered response to that criticism rather than a refutation of it.
What to watch next: independent replications of the benchmark numbers, any reproduction of the "geometry stays in after the teacher leaves" claim, and whether the same training recipe transfers to robots with more varied sensors or longer task horizons. The next preprint in this thread will tell us whether the gain is a property of the recipe or a property of these two benchmarks.