Robotics papers — 2026-10-18

The Morning Preprint — this day as one edition: the day's papers, the five things worth knowing, and the episodes behind them.

This is a look at making robot vision language action models more connected to real physical space, which is important because these models often use 2D image coordinates instead of actual camera geometry. This problem becomes a big issue when systems use several cameras where the relationship between the views is known but not used in the visual processing.

The most important work involves G cubed VLA, which adds a camera-aware geometric module to inject calibrated structure into the visual-token stream of pretrained VLA models without changing their original action space or learning objective. This is done by combining intrinsic-conditioned ray embeddings, projective positional encoding, and bidirectional cross-view fusion to create geometrically consistent representations across different camera views.

This geometric grounding is useful because it allows training using ground truth point maps or confidence-gated teacher predictions, which has shown improvements on tasks like RoboTwin2.0 when using three cameras with ground truth supervision. This consistency is also tested in camera-shifted scenarios, where the gains depend on correct camera geometry but remain stable under mild calibration changes.

Another important direction focuses on structuring latent actions to better capture how much change a robot makes and what kind of change it makes, which is addressed by PoLAR. PoLAR imposes a radial structure on latent actions, using temporal offsets as a proxy for transition extent to encourage larger radii for larger changes in the action space. This approach has been shown to improve downstream policy performance in both simulation and real-world robot experiments compared to existing latent action baselines.

For navigation in uncertain environments, Schrödinger's Navigator proposes a belief-aware framework that reasons over multiple imagined 3D futures at inference time, maintaining a superposition of plausible scene realizations instead of committing to one map. This method uses an adaptive occluder-aware sampler to focus imagination on uncertain regions and aggregates these futures into a Future-Aware Value Map for safer action selection.

Finally, there is work tackling planning under uncertainty in Partially Observable Markov Decision Processes using VOiLA, which learns task-agnostic world models through diffusion models to enable efficient online planning. This framework has shown that the distillation strategy can reduce sampling costs by nearly three orders of magnitude while achieving performance equal to or better than current reinforcement learning methods for POMDPs.

The most important takeaway is that allocating test-time compute intelligently, rather than scaling it uniformly, is key to achieving high performance for embodied agents where deployment matters most. This idea is central to DIRECT, which introduces a routing framework that uses multimodal scene context to allocate compute at each step of a closed-loop agent by selecting a planner for every subtask.

This routing approach shows that test-time compute is not a simple lever; different scaling axes, like chain-of-thought depth or model size, lead to qualitatively different gains. For instance, experiments on VLABench and RoboMME demonstrated that this selective allocation can match or exceed the success rate of a stronger model while using up to sixty five percent less average latency on a physical Franka arm in a DROID setup.

Furthermore, this idea connects to how policies handle memory, as RMBench systematically evaluates manipulation tasks spanning multiple levels of memory complexity, highlighting where existing policies struggle with reasoning over historical observations. This systematic evaluation helps inform the design of modular policies like Mem-0, which explicitly include memory components for controlled study.

Today's papers

The papers

Important terms

G cubed VLA
This work adds a camera-aware geometric module to vision language action models. It injects calibrated structure into visual tokens using ray embeddings and cross-view fusion, making the model aware of real physical space.
PoLAR
This technique structures latent actions radially. It uses temporal offsets to represent how much change a robot makes, encouraging larger changes for larger action radii in simulation and real tests.
Schrödinger's Navigator
This framework plans in uncertain environments by reasoning over multiple imagined 3D futures at inference time. It maintains a mix of possible scenes instead of committing to just one map.
DIRECT routing framework
This introduces a way to intelligently allocate test-time compute. It uses scene context to select the best planner for each small task, showing that selective allocation is more effective than uniform scaling.