Robotics papers — 2026-10-20

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

Today's work centers on how robots can use tools in novel ways, which is crucial because tool use dramatically expands what robots can accomplish beyond simple preprogrammed actions. This is addressed by GROW squared to tackle the challenge of open-world affordance grounding, which means figuring out which object to use as a tool and precisely where on that object to apply force. This method splits the process into semantic understanding using vision-language models and geometric localization using vision foundation models, bypassing heavy training requirements.

This is significant because it allows robots to use any open-category object as a tool, not just those explicitly shown in training data. This capability is demonstrated by its success in zero-shot generalization over open-category objects and outperforming baselines in both simulation and real experiments. This grounding process connects to VLA-Pro, which aims to improve cross-task generalization in vision language action models by storing task-specific procedural memories that can be retrieved during inference.

VLA-Pro improves on existing models by using LoRA adapters as procedural memories, allowing them to transfer experience between different tasks and significantly boosting real-world success rates from five point eight percent up to sixty-five point zero percent in manipulation tasks. This memory transfer mechanism is complementary to GROW squared's focus on grounding the tool itself.

Furthermore, we are exploring how robots can execute complex sequences reliably using TRACT, which introduces temporally routed action chunks and chronological phase authority to shorten the decision horizon for imitation learning. TRACT achieves a full-sequence success rate of ten out of ten across six real-robot variants, showing that structuring temporal information helps eliminate phase ambiguity and reduces stalls.

Finally, GRID-Planner is a graph-structured relational reward shaping framework for diffusion planning used in autonomous driving to improve trajectory selection under complex interactions. By combining graph importance weighting with a time-indexed relational reward surrogate, GRID-Planner reduced collision rates by thirty five point eight percent on nuPlan and improved the planning score by three point four percent.

The hierarchical optimisation framework for integrated electric hydrogen transport systems is most significant because it tackles the complex operational challenges arising from coupling vehicle scheduling with multienergy dispatch in growing electric and hydrogen transport sectors. This approach uses a solver-free greedy heuristic to handle vehicle scheduling, which then feeds demands into a deep reinforcement learning approach for energy dispatch, aiming to minimize total operational cost while meeting scheduled transport needs.

The deep reinforcement learning approach is key because it learns how to allocate battery operation, hydrogen-tank operation, and solar power generation optimally without needing retraining when faced with different transport demands or solar penetration levels. This learned policy shows strong generalisation capability for practical deployment in real-world scenarios.

ToPos offers a way to improve spatial sampling in challenging terrains by treating the terrain as a manifold and using geodesic distances instead of simple flat-map distances, resulting in about a seventy-four percent improvement in optimal surface area-balanced distribution. This method mitigates planimetric distortion by respecting the actual surface geometry.

The control co-design framework addresses sustainability issues by categorizing environmental metrics across manufacturing, operation, and disposal stages to support trade-off analysis for energy systems like microgrid data centers. This framework shows substantial improvements in environmental sustainability categories compared to baseline procedures.

Reduced arc-weight heterogeneity for master stability analysis shows that one coupling strength is often sufficient to achieve a diagonalizable Laplacian matrix in directed networks, simplifying the analysis of synchronization modes. This finding reduces coupling heterogeneity while retaining the simpler modal form of master stability analysis.

LLM-Flock proposes a decentralized multi-robot flocking framework that uses influence-based consensus and collision avoidance to improve formation stability when using large language models for decision making. This method demonstrates notable improvements in convergence over previous LLM-based approaches in physical drone systems.

Today's papers

The papers

Important terms

GROW squared
A method for grounding tool use by combining semantic understanding from vision-language models and geometric localization using vision foundation models to figure out which object to use as a tool.
VLA-Pro
An improvement to vision language action models that uses LoRA adapters as procedural memories, allowing the model to transfer learned experiences between different tasks for better generalization.
TRACT
A framework for executing complex sequences reliably in robots by introducing temporally routed action chunks and chronological phase authority to reduce decision-making stalls.
GRID-Planner
A graph-structured reward shaping framework used in diffusion planning for autonomous driving, which uses relational rewards to improve trajectory selection under complex interactions.
Hierarchical optimisation framework
A system for integrated electric hydrogen transport that uses a solver-free heuristic for scheduling and deep reinforcement learning for energy dispatch to minimize operational costs.