Robotics papers — 2026-10-04

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

Today we are diving into how we can make complex control problems physically possible. This is crucial because standard optimization methods often get stuck in infeasible solutions due to actuator limits or strict deadlines. The work that matters most here is the geometric decision procedure for STL feasibility and repair, which tackles this by evaluating physical feasibility without needing to discretize the time horizon.

This method works by transforming temporal logic constraints into continuous spatial backward reachable sets. It then uses an analytical inversion to map temporal windows into exact, closed-form temporal delays. This effectively fixes boundary deficits to restore realizability.

We also have work on improving how we handle failures in hardware, specifically vulnerability-weighted routing for timing-critical nets in SRAM-based FPGAs. This methodology incorporates predicted routing-fault severity directly into the objective function to selectively reroute only the highest-risk nets. This has shown a significant reduction in configuration-induced timing vulnerability while maintaining nominal performance.

This is complemented by research on improving imitation learning policies where Anchor-Align augments behavior cloning by distilling layer-wise representations from a frozen vision language model. This prevents representation drift and aligns language and action predictions.

Finally, we look at how to make autonomous navigation safer in multi-agent systems with Risk-Bounded Multi-Agent Visual Navigation via Iterative Risk Allocation. This dynamically distributes risk budgets among agents during pathfinding to allow for more efficient travel through hazardous areas.

The most significant work here is HumanoidTTT because it tackles a real problem in deploying humanoid robots where reusing validated motions efficiently is crucial for continual operation. This framework introduces Selective Full-Motion Reuse, which allows direct reuse of complete motions only from certified starting points. This bypasses the need to generate new ones when the entry state is known.

This mechanism, combined with Test-Time Capability Consolidation which adapts which capabilities are kept in a store based on future use feedback, has shown zero unsafe accepts and a sixteen point four times speedup over fresh generation.

This efficiency relates closely to how we manage control policies in other domains. For instance, the paper on Model-Free Output Feedback Stabilization via Policy Gradient Methods explores stabilizing unknown linear dynamical systems using policy gradient methods. This is a foundational control problem.

Furthermore, the work by Diffusion-Guided Multi-Arm Motion Planning addresses scalability in complex planning tasks by decomposing them into single-agent problems coupled with collision resolution. This mirrors the need for careful state management seen in motion reuse.

We also see related concerns when considering thermal constraints. ThermE manages shared thermal headroom for LLM inference on edge SoCs by predicting and jointly managing this resource. It shows a forty point five five percent reduction in time-to-first-token relative to vLLM.

This predictive management contrasts with the reactive nature of some control methods, such as those explored in Interaction Dynamics Modeling and Predictive Control for Safe Steerable Catheter--Tissue Interaction. These focus on regulating complex catheter--tissue interaction dynamics through impedance control or other mechanisms.

Finally, there is a concern about fidelity when using simulations for validation. Bounded-Fidelity Sim-as-Demo-Stage proposes suppressing contact physics during handoffs to ensure audit chain reproducibility in governance benchmarks. This approach maintains near zero divergence in event logs compared to methods that integrate full contact dynamics.

The most significant development here is the work by eVGGT because it directly tackles the performance-latency bottleneck in geometry-aware vision models. This is critical for deploying robust visuomotor policies in real robotics. This paper introduces eVGGT, a lightweight version distilled from VGGT that offers a substantial speedup and memory reduction while maintaining strong task performance.

This efficiency gain is complemented by the work on Task and motion planning by Shcherba Eckart Cobo-Briesewitz. They address the interface between high-level symbolic plans and low-level continuous trajectory generation. Their method uses meta-optimization to bridge this gap, employing program search over trajectory optimization problems as an interface between a foundation model and robot control.

This is important because it moves beyond simple skill chaining or direct joint prediction by creating a structured way for the system to reason over both planning and execution.

A related concern in complex manipulation is addressed by MILE, which presents a teleoperation-based data collection system using an exoskeleton and robotic hand. MILE showed high success rates during teleoperation, especially when paired with tactile input. This suggests that incorporating rich sensory feedback significantly improves imitation learning performance for dexterous tasks.

This connects to the vision encoder work because better sensory input is often needed to train effective policies like those mentioned in eVGGT.

Furthermore, legibility in planning is being explored by a method that extends previous research on PDDL domains to create plans that are more easily understood by an observer. This work shows that plan legibility involves a trade-off with efficiency. We have to balance how clear the plan is against how fast it can be executed.

This concept of structured planning interfaces with the need for efficient control suggested by the path tracking controller developed for agricultural tractors using Nonlinear Model Predictive Control. This controller achieved a mean absolute cross-track error of 6.1 cm during field testing.

The most significant finding relates to developing a new method for safe multi-agent navigation guided by goal conditions. This is crucial for autonomous systems operating in dangerous settings. This approach integrates goal-conditioned reinforcement learning with safe reinforcement learning to learn policies while simultaneously estimating safety levels using a self-training algorithm that prunes unsafe edges in a graph and generates waypoint plans.

This method extends to multiple agents by leveraging Conflict-Based Search to create waypoint-based plans, showing effectiveness against state-of-the-art baselines in complex environments.

This work builds upon the idea of planning but addresses the limitations of traditional methods that rely on fixed distance metrics. It combines the strengths of goal-conditioned RL and safe RL to learn navigation policies without manual heuristics. This is vital for scalability in multi-agent scenarios.

This advancement complements the work on Model Predictive Path Integral control, which uses sampling to find optimal controls while Control Barrier Functions enforce safety constraints by augmenting the system state with a parameter whose derivative becomes an additional control input optimized by MPPI.

The development of an EMT model for the Balearic power system is also important because it addresses stability challenges arising from higher integration of inverter-based resources in island power systems. These systems are becoming more susceptible to disturbances, and this model is necessary for analyzing the stability due to the high amount of power-electronics devices present in that system.

Furthermore, a new VSC-HVDC interconnection between the Iberian Peninsula and the Balearic Archipelago is planned as a key enabler for energy transition in those islands. This link will use Voltage Source Converter technology and consist of a bipole with 2x200 MW capacity each, facilitating the decarbonization efforts there.

Finally, the identification of steering and speed control systems for a four-wheel-steering agricultural tractor is relevant. This provides a data-driven model combining kinematic models and estimated actuator dynamics to develop path tracking control. This work tackles the challenge of modeling complex digital and mechatronic systems using first-order actuator models with estimated time constants.

Today's papers

The papers

Important terms

Geometric Decision Procedure for STL Feasibility and Repair
This method checks if a complex control problem is physically possible by turning temporal logic constraints into continuous spatial sets. It finds exact time delays to fix feasibility issues without needing to discretize time.
Vulnerability-Weighted Routing
This technique improves hardware reliability in FPGAs by incorporating predicted routing fault severity directly into the optimization goal. It selectively reroutes only the most risky connections for better timing.
Anchor-Align
This is a method to improve imitation learning by combining behavior cloning with distilling representations from a frozen vision language model. This prevents representation drift between language and action predictions.
HumanoidTTT / Selective Full-Motion Reuse
This framework allows humanoid robots to reuse validated motions efficiently by only reusing complete motions from certified starting points. This significantly speeds up motion generation.
eVGGT
This is a lightweight, distilled version of a geometry-aware vision model. It offers substantial speed and memory gains while maintaining strong performance for real robotics applications.