Robotics papers — 2026-10-09

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

This is how robots can feel the world through touch, which is important because current vision language action models mostly ignore contact information for manipulation. The main idea is that if you show the contact signals in the same visual frame as the scene, any image-conditioned policy can use them right away without needing a complete architectural change.

Testing this concept with Visible Touch used a custom low-cost magnetic sensor made from off-the-shelf parts, and this approach showed significant gains across benchmarks. Specifically, Visible Touch improved BC-Transformer success by fifteen point seven percentage points on average in the two view setting, and similar improvements were seen in the one view setting. When fine-tuning pretrained vision language action models, miniVLA gained twenty-five percentage points on average on LIBERO tasks, and for four real world contact rich tasks, it achieved a thirty point zero gain with our custom sensor.

Integrating this tactile data is important because it shows how directly usable contact signals can be when they are spatially aligned with visual input. This contrasts with other work that focuses on recovery; for instance, Failure-Aware Retry proposes a framework called FAR that lets robots learn from failures at test time by combining failure contrastive preference adaptation with lightweight action perturbations during retries.

Furthermore, efforts are being made to make these complex tasks more reliable in the real world. RoboFAC introduces a framework that creates a large dataset of erroneous manipulation trajectories and uses it to build a multimodal model for failure analysis and correction, achieving thirty-four point one percent higher failure analysis accuracy compared to GPT-4o. This suggests that systematic diagnosis can significantly boost recovery capabilities.

We are also exploring how robots can handle long horizon tasks by distilling human feedback into reusable skills through LYRA, which uses external memory and retrieval augmented generation guided by user hints to extend learned behaviors across many interactions.

The work that matters most is how we can move beyond static goal representations when dealing with multi-goal robotic tasks, because this addresses a fundamental weakness in current planning methods where conflicts lead to local minima AKMPC. This paper shows that failures often come from not modeling how objectives trade off based on the changing world state, suggesting that reactive control can be made robust by structuring the interaction graph differently.

This idea is supported by other work focusing on how robots interact with ambiguous instructions, such as SAIN, which tackles the problem of robots needing to ask clarifying questions in dialogue rather than just accepting text hints for navigation. SAIN converts these interactive dialogues into persistent spatial and object-centric memories that a unified policy can use for frontier ranking, showing how dialogue can be turned into actionable state information.

We also see efforts to handle complex sequential tasks by unifying world models with execution policies; WorldDP introduces a hierarchical framework where a high-level world model guides the low-level diffusion policy to solve multi-stage manipulation problems better than previous single-stage methods. This is complemented by FLASH Policy, which tackles the latency issues in learning visuomotor policies by using sparse history anchoring and Legendre polynomials to achieve very fast inference times, demonstrating state of the art success rates on manipulation tasks.

On a different scale, there is work addressing physical design challenges; SurGE uses surrogate gradients within CMA-ES to accelerate the co-design of legged robots with elastic elements, showing significant improvements in population concentration and objective reduction when compared to standard evolutionary algorithms. This contrasts with control system optimization, where VSC-HVDC setpoint adjustment for maximum grid utilisation under voltage constraints provides an analytical method for adjusting power setpoints in real time to maximize loadability under voltage stress using local measurements.

Finally, when considering high-speed autonomous driving, OptCar presents a recipe for specializing generalist vehicle models by conditioning them on state history tokens to bridge the gap between generalist models and specialist performance across different terrains. This approach shows that targeted specialization can yield substantial gains in tracking error when dealing with complex dynamics like slip dominance.

The most significant development today concerns how to unify complex motor skills into a single controllable policy because current methods often treat grasping, relocating, rotating, and translating as separate problems that don't communicate well during long tasks. This is addressed by UniCross which introduces a unified framework modeling all four skills within the same state and action spaces, allowing for the distillation of one cross-skill policy that handles unseen objects and remains robust to disturbances.

This idea connects to how agents can control robots directly through visual interfaces because VIA shows that foundation models already possess the necessary competence for robot control when given a proper interface, essentially suggesting that a coding agent is secretly a robot controller. Furthermore, the ability of policies to adapt to new sensory inputs is crucial, as Multisensory Continual Learning demonstrates that augmenting vision-only models with limited multisensory data can improve general capabilities beyond what they learned during initial training.

This focus on policy steering also touches upon how human expertise can guide autonomous systems; SAPS shows that blending real-time human teleoperation commands with pretrained vision language action models creates a lightweight framework that achieves high hardware success while significantly reducing the cognitive load on the human operator. The underlying concept of shared control is also relevant to how different agents might interact in larger settings, as the taxonomy of collaboration in multi-robot systems helps clarify the distinctions between cooperation and coordination when tasks become too complex for a single entity.

NovaPlan is the most significant piece of work here because it tackles the core challenge of complex long-horizon manipulation by proposing a hierarchical framework that systematically plans, verifies, and repairs visual plans. This matters because combining vision-language models with video generation often leads to errors in physics or geometry when trying to execute real tasks.

The framework works at the high level by having a VLM planner decompose the task and check several potential video rollouts to filter out futures that look inconsistent. This is then translated into physical actions using a hybrid geometric representation that can switch between object-centric flow and human hand flow depending on what is needed. Finally, it closes the loop by constantly watching how things are executing and synthesizing small corrective behaviors, like fingertip poking, when mistakes happen.

This approach builds upon earlier work where imitation learning is used to teach robots tasks; specifically, Responsive Noise-Relaying Diffusion Policy (RNR-DP) improves on Diffusion Policy by using a noise-relaying buffer to ensure actions are responsive to the latest observations. RNR-DP achieves an eighteen percent improvement in success rate on response-sensitive tasks compared to Diffusion Policy.

The concept of agentic scene policies also addresses generalization, suggesting that unifying grounding and robot skills through a scene-agent tool interface can allow policies to work across different objects and instructions zero-shot. This contrasts with modular policies that often struggle with explicit spatial reasoning when retrieving objects based only on semantics.

While not directly related to manipulation, the work on Global Geolocated Realtime Data of Interfleet Urban Transit Bus Idling provides a massive dataset for understanding large-scale operational patterns, detecting about two hundred thousand idling events daily across numerous cities. This data is crucial for urban planners looking to reduce emissions and improve traffic management.

Stability analysis in multi-constraint safety filters shows that while control barrier functions guarantee forward invariance, they can sometimes lead to unstable active-set modes, requiring geometric frameworks to characterize when divergence occurs versus bounded behavior. This informs how we design safer systems for constrained linear dynamics.

The work by Where Success Breaks: Failure-Boundary Learning for Robust Vision-Language-Action Models is the most significant because it tackles the fundamental problem of making vision language action models robust when they fail in unpredictable ways, moving beyond simple supervised fine-tuning. This approach suggests that instead of just learning what works, we should learn where the success breaks and shape the boundary around that failure.

This idea is supported by Instrumentation for Imitation Learning: Enhancing Training Datasets for Clothes Hanger Insertion, which shows that adding sensor data to imitation learning significantly boosts performance, with policies leveraging this instrumentation outperforming vision-only counterparts by fourteen to twenty-five percent points. Furthermore, Dynamic Neural Koopman Distillation for Fast Robot Control Using Diffusion Models addresses the latency issue in diffusion models for robot control by distilling multistep inference into a single forward pass, achieving competitive performance with reduced inference parameters on long-horizon tasks.

The theoretical underpinning of identification limits is provided by Privacy-Preserving Cram'er-Rao Lower Bound, which establishes the fundamental limit for identification accuracy when privacy constraints are applied using Fisher information as the metric. This work shows that precise privacy-preserving bounds can be established explicitly, and recursive formulas are developed to handle multi-measurement systems efficiently.

Finally, there is the connection between control stability and learning frameworks in Learning Robust Control Lyapunov Functions through Lipschitz Neural Networks, which introduces a framework for jointly learning Lyapunov functions and controllers for nonlinear systems subject to disturbances. This work uses higher-order bounds to accelerate verification, linking the need for robust learning with the structural constraints of neural networks in control theory.

The most significant development is the TriDeliver framework, which introduces a hierarchical cooperative structure integrating human couriers, UAVs, and ground vehicles to achieve efficient instant delivery. This matters because it tackles the inherent limitations of using any single agent alone by showing a substantial reduction in delivery cost versus existing state-of-the-art methods. The paper demonstrates that by leveraging transfer learning to move knowledge from human couriers to UAVs and ground vehicles, TriDeliver cuts the overall delivery cost by 65.8% while simultaneously improving delivery time and impacting the original tasks of the ground vehicles significantly.

A related but less immediately impactful contribution is the work on confidence-aware safe and stable control for control-affine systems AKMPC. This research addresses synthesizing safe and stable control inputs for nonlinear systems when full state measurements are missing by adapting control Lyapunov function and control barrier function techniques to an output feedback setting. The simulation studies show that this approach leads to improvements in both the observer's estimation accuracy and the fulfillment of safety requirements, which connects it conceptually to how other systems must maintain reliable states.

The work on collision avoidance for convex primitives via differentiable optimization based high-order control barrier functions Collision Avoidance for Convex Primitives via Differentiable Optimization Based High-Order Control Barrier Functions tackles the critical issue of ensuring safety in dynamic systems, specifically collision avoidance among convex shapes. This paper introduces a high-order control barrier function framework that transforms nonconvex safety constraints into linear ones using differentiable optimization, which allows for torque control and includes a circulation mechanism to prevent undesired equilibria.

Finally, the study on self-mixing laser interferometry for robotic tactile sensing Self-Mixing Laser Interferometry for Robotic Tactile Sensing presents a novel way to sense slip and extrinsic contact using microvibrations from a robotic fingertip. This method was found to be more sensitive to subtle slip events and more resilient against noise compared to acoustic sensing, offering a promising new branch for tactile sensing in robotics.

Today's papers

The papers

Important terms

Contact Signals
These are tactile signals from a robot's touch that are shown alongside visual frames. They are crucial because current vision models often ignore this contact information when manipulating objects.
Vision Language Action Models (VLAMs)
These are AI models that can understand both what they see and what actions to take. The research focuses on how to integrate touch data into these models without needing a complete architectural overhaul.
Failure-Aware Retry (FAR)
This is a framework that helps robots learn from mistakes during testing by combining failure contrastive preference adaptation with small action changes when trying again.
NovaPlan
This is a hierarchical planning framework for complex manipulation. It uses a visual language model to plan, checks potential video outcomes for consistency, and corrects errors in real-time.