Robotics papers — 2026-10-23

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 embodied intelligence safer by improving reinforcement learning for vision language action models. The most important work is SafeDojo, which proposes a model-based safe reinforcement learning framework that uses world-model imagination to improve policies online on top of an interactive video world model SafeDojo. This approach is significant because existing safe methods either demand costly real-world exploration or rely on manually designed safety functions, neither of which scales well for vision language action models in open environments.

The core mechanism involves decoupling task execution from trajectory safety by using a ResNet success classifier to score future frames and a lightweight safety head that predicts per-step contact probabilities from the current latent context and proposed action chunk, all balanced through Lagrangian-constrained GRPO. This framework showed promising results, improving the average safe-success rate by 8.25 percentage points on SafeLIBERO Level I and remaining competitive with strong model-free baselines in zero-shot Level II settings.

Another important piece of work is mu zero which introduces a scalable world model based on 3D traces to enable robot learning without relying on embodiment-specific action labels mu 0. This model forecasts smooth 3D trajectories for salient interaction points, and its TraceExtract system pretrains it by combining a vision-language backbone with a modular trace expert to extract 3D supervision from video sources.

For complex manipulation tasks, Multi-Resolution Tactile Imitation Learning (MiTaS) explores fusing multiple heterogeneous tactile sensors operating at different temporal resolutions to solve contact-rich manipulation problems MiTaS. This framework uses modality-specific convolutional stems and transformer-based fusion to integrate information from RGB cameras, a GelSight Mini sensor, and an Evetac event-based sensor, achieving an average success rate of 80% on five contact-rich manipulation tasks.

Finally, there is work on agentic skill programming through iterative robot exploration called ASPIRE ASPIRE. This system autonomously writes and refines robot control programs in a code-as-policy paradigm, using a closed-loop execution engine to diagnose failures and distill validated fixes into a reusable skill library. ASPIRE has shown significant gains, surpassing prior methods by up to 77% on manipulation tasks under perturbation and achieving zero-shot generalization on representative unseen long-horizon tasks.

The most pressing work here is about creating reliable forecasting tools for prosumer flexibility in ancillary service markets because without them, demand response aggregators cannot accurately bid into frequency controlled services where reliability standards like the P90 are very strict. This paper proposes a scalable uncertainty quantification framework that combines Monte Carlo dropout with conformal prediction to give calibrated prediction intervals for aggregated prosumer flexibility.

This approach is significant because conventional methods often fail; specifically, standalone Monte Carlo dropout methods systematically overestimate available flexibility and violate P90 compliance, which means they suggest more power availability than is actually there. The proposed MCD-CP framework, however, achieves reliable coverage with controlled conservatism when embedded in the aggregator bidding model, substantially reducing overbidding risk and reaching up to seventy percent of perfect-information profit while meeting regulatory constraints.

The work on action-conditioned world models for robotic ultrasound is also important because it tackles the difficulty of training autonomous tasks where high-quality demonstrations are hard to get and explicit simulators are complex due to view-dependent artifacts. This pipeline uses a latent conditional diffusion world model to predict future observations, which then feeds into a goal-conditioned temporal transformer that predicts ordered probe motions based on rewards from the frozen world model.

Another area of relevance is how data center growth impacts grid resilience, as the rapid scaling of AI workloads concentrates demand and creates new risks under disruptive events. This paper extends a multi-time-step DC optimal power flow framework to evaluate unserved energy when aggregated data center demand is added to an IEEE 30-bus system. The results show that temporally concentrated data center demand can amplify resilience impacts, increasing total unserved energy significantly even without increasing the total energy consumption of the system.

Finally, there is research into decentralized discrete-event systems where comparing different local decision fusion rules is difficult because existing methods lack a uniform way to compare architectures. This paper extracts a method based on morphisms between graphs to provide a uniform approach for comparing the permissiveness of these decentralized architectures.

The most significant development this week involves creating a framework for coordinating traffic light systems and connected automated vehicles in mixed-traffic intersections because it directly addresses the practical challenge of safely integrating autonomous vehicles into existing urban infrastructure. The work by Viet-Anh Le Andreas A. Malikopoulos, titled Distributed Mixed-Integer Quadratic Programming for Mixed-Traffic Intersection Control, develops a distributed solution framework using a mixed-integer quadratic program to optimize traffic light coordination with connected automated vehicles and traffic lights under high penetration rates. This approach uses proximal ADMM to solve the convex relaxation of the MIQP while employing a finite-grid tightening rule to maintain integrality requirements, which is crucial for ensuring real-world feasibility.

This optimization work builds upon earlier efforts in control theory by addressing stability in systems with limited regions of attraction. For instance, Lorenzo Fici Dalim Wahby Alvaro Detailleur Matthieu Barreau Guillaume Ducard's study on Region of Attraction Estimation for Linear Quadratic Regulator, Linear and Robust Model Predictive Control on a Two-Wheeled Inverted Pendulum investigates how to estimate the region where a two-wheeled inverted pendulum system will converge using both analytical bounds and Monte Carlo estimation. This provides a practical way to evaluate controller performance beyond overly conservative analytical guarantees.

Another area of research focuses on improving robotic decision-making by refining how models handle real-world discrepancies. Tuo An et al.'s Feedback World Model Enables Precise Guidance of Diffusion Policy proposes an inference-time mechanism that closes the loop between prediction and observation by maintaining a lightweight feedback state to correct future predictions using real-time data, which has shown substantial improvements in prediction accuracy under distribution shift. This contrasts with static world modeling by allowing the system to adapt online.

Furthermore, efforts are being made to make robotic learning more structured and interpretable for complex tasks. Yiyuan Pan et al.'s Emergent Neural Automaton Policies introduces a framework that allows a bi-level neuro-symbolic policy to emerge from visuomotor demonstrations by first inferring a Mealy state machine to capture latent task modes, which then guides a low-level network for precise control. This structured approach shows promise in achieving high sample efficiency for long-horizon tasks.

In the realm of human demonstration utilization, Harry Freeman Chung Hee Kim George Kantor's WARPED framework synthesizes realistic wrist-view observations from monocular RGB data to train visuomotor policies, requiring significantly less data collection time than previous methods. This synthesis is then used to directly train a robotic policy.

Finally, there are advancements in perception for safety-critical applications. Dian Yu Qingchuan Zhou Bingkun Huang Majid Khadiv Zewen Yang's Safe-Night VLA proposes integrating long-wave infrared thermal perception into vision-language models to enable robots to see unseen areas while enforcing safety constraints through control barrier functions during execution. This demonstrates how foundation models can leverage non-visible physical modalities for robust manipulation.

The most significant finding relates to how hierarchical decomposition can dramatically speed up exploration in complex settings, which is central to HEHA HEHA. This method tackles the massive constrained multiple hamiltonian path problem by breaking it down into global and local planning stages, and its routing algorithm RCH manages this decomposition effectively to find solutions that are up to forty-seven percent faster in terms of makespan compared to existing methods HEHA.

This efficiency gain is supported by the work on robot manipulation, specifically AFFORD2ACT AFFORD2ACT, which shows that using only sparse keypoints for policy input can still retain strong manipulation capability across various tasks, achieving eighty-two percent average success in held-out settings. This contrasts with the limitations seen when replacing learned gating with uniform weights, which caused success rates to drop significantly to fifty-two percent AFFORD2ACT.

Furthermore, the ability to navigate complex instructions is being addressed by AgriVLN AgriVLN, which introduces an STL module to break down long natural language instructions into simple subtasks, improving performance on short instructions while slightly increasing error on long ones. This contrasts with the observation that existing vision-and-language navigation methods perform poorly on longer commands AgriVLN.

On a more foundational level, safety in control systems is being explored through stochastic neural control barrier functions Stochastic Neural Control Barrier Functions, which proposes frameworks for synthesizing these functions under different activation types and validation settings for smooth and ReLU cases. This work builds upon the need for robust control, similar to how DEFT DEFT models branched objects in real-time using differentiable physics and learning, demonstrating accurate prediction for wire-harness assembly tasks.

Today's papers

The papers

Important terms

SafeDojo
A model-based safe reinforcement learning framework that uses world-model imagination to improve policies online, avoiding costly real-world exploration or manual safety functions for vision language action models.
mu zero
A scalable world model based on 3D traces that enables robot learning without needing specific embodiment labels. It forecasts smooth 3D trajectories for interaction points using a vision-language backbone.
MiTaS
Multi-Resolution Tactile Imitation Learning, which fuses heterogeneous tactile sensors (RGB, GelSight Mini, event-based) using modality-specific stems and transformers to solve contact-rich manipulation problems.
ASPIRE
A system for agentic skill programming that autonomously writes and refines robot control programs in a code-as-policy paradigm, distilling validated fixes into reusable skill libraries.