Robotics papers — 2026-10-19

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

The work that matters most is KODA because it converts egocentric human demonstrations into whole-body vision language action training data using a reusable canonical trajectory to guide robot view synthesis and generate controller specific supervision. This is important because training whole body policies demands aligned robot views, and KODA solves the problem of misaligned views in human demonstrations by synthesizing them from a canonical record.

This process involves first generating a simulator checked whole body trajectory using sparse trackers and sampling based command optimization, which then dictates the camera poses for synthesizing robot view observations. For each deployment controller, KODA optimizes commands to track this stored motion, recording action labels and proprioceptive states; when switching controllers it regenerates supervision for co training with matching demonstrations.

A related piece is MOSS which addresses how generative policies select modes when multiple actions are feasible in manipulation tasks by turning mode selection into an explicit choice of where generation starts. This method constructs semantically aligned mode specific Gaussian sources that encode mode identity in the initial noise, and a modality strength estimator adaptively controls the geometry of these sources to steer to the target behavior with high success rates on real robots.

We also have CQDC-Warp which presents a GPU accelerated differentiable quasi dynamic simulator for contact rich manipulation. This simulator uses gradient based planning to jointly optimize motion and contact decisions by specifying intended contacts before planning and encouraging them through soft costs. This framework achieves high success rates on in hand rotation benchmarks while significantly reducing trajectory generation time compared to existing methods.

For long horizon control, Mem4Act introduces a framework for learning effective latent memory as a fixed size representation explicitly optimized for long horizon robotic manipulation. It organizes history via multi granularity temporal aggregation and uses question answering supervision to ensure the policy actually relies on memory, resulting in state of the art performance across various manipulation benchmarks by capturing both high level task state and fine grained perceptual evidence.

Finally, we have when to ask for help which provides a decision theoretic account for slow module scheduling in dual system vision language action models. This formalizes invocation as a calibrated decision problem whose fixed ratio, fixed budget, and history gated schedules are degenerate cases of the value of intervention; this work turns slow module invocation from a design opinion into a measured decision by testing adaptive gains against cost ratios across orders of magnitude.

The most significant work here is LIN-SLAM because it introduces a neural implicit backend to LiDAR inertial SLAM that achieves real-time online operation on mobile robots while maintaining top accuracy, which is a major step forward for practical navigation. This method couples high-rate geometric LiDAR inertial estimation with neural implicit tracking, continual mapping, and global refinement to overcome the limitations of existing methods.

This approach is important because it solves the problem of achieving state-of-the-art localization accuracy without the prohibitive computational cost associated with purely neural implicit SLAM techniques. The experiments show that LIN-SLAM achieves the lowest overall mean absolute trajectory error among state-of-the art geometric and neural implicit methods, while significantly reducing runtime compared to other neural implicit SLAM methods.

The underlying mechanism involves coupling high-rate geometric LiDAR inertial estimation with neural implicit tracking, which is then refined through continual mapping and global refinement; this is more important than the specific hardware setup because it establishes a novel hybrid architecture for real-time onboard SLAM.

Further supporting this direction, MoCaP presents an agentic framework that turns passive human video recordings into guidance for end-to-end mobile manipulation policies. This work is relevant because it shows how to leverage unstructured human experience to solve complex manipulation tasks without requiring explicit demonstrations for every scenario.

The success of MoCaP is highlighted by its ability to improve object navigation success from seventy percent to one hundred percent and raise the rate of finding a reachable base pose from zero percent to one hundred percent at an unfavorable start location, contrasting with existing approaches that cannot succeed at these tasks.

Another piece of work, Fast-VLA, addresses the issue of high computational demands in vision-language-action policies by introducing a mechanism where a frozen foundation model is used once per episode to establish fixed task memory. This is important because it allows the system to predict action chunks without repeatedly recomputing the task context whenever it predicts a new action.

Fast-VLA achieves substantial reductions in amortized policy inference time, cutting it by ten point nine to seventeen point one times and requiring only a small amount of post-prefill online resident memory; this demonstrates that task context can be established once while the control loop remains responsive throughout execution.

TS-WAM is also relevant as it addresses the limitation of relying on action-labeled robot demonstrations by introducing self-supervised temporal learning to predict future visual states. This framework learns from heterogeneous data, including both robot demonstrations and action-free human videos, to guide action generation at inference time; this is important because it aims to improve generalization and robustness in manipulation tasks by explicitly predicting future visual features that guide the policy. TS-WAM achieves strong manipulation performance across various benchmarks while maintaining a lightweight deployed policy.

Finally, Compliant GMT proposes a modular framework for full-body compliant responses in general motion tracking, addressing the safety concerns associated with stiff policies in humanoid robots. This work is significant because it combines the generality of pretrained general motion trackers with a residual adapter that introduces compliance conditioned on proprioception history.

The work that matters most right now is the development of the framework for Inertially Significant Humanoid Teleoperation via Whole Body Model Predictive Control and Online Parameter Estimation because it directly tackles the critical need for humanoid robots to sense and adapt to changing inertial parameters when handling significant payloads, which is essential for practical applications. This paper shows how integrating a whole body model predictive controller with an extended Kalman filter allows the system to estimate time-varying inertial parameters online, feeding these estimates back into the controller for more accurate feedforward torques under external load.

This approach builds upon other efforts in control and manipulation. For instance, CLIS-MPPI addresses issues with planning delay in motion control by modeling execution-time state mismatch as a distribution over initial states, which has shown improvements in safety and path consistency on a 6-DoF UR5e manipulator compared to standard MPPI. Furthermore, the work on Spark-0 shows how tactile sensing can serve as scalable supervision for dexterous visuomotor pretraining; by using touch features alongside visual tokens during pretraining, Spark-0 achieved higher success rates on both simulated and real-robot tasks compared to vision-only methods.

Another key area involves understanding robot capability. The Attribution-Gated Joint Estimation of Intention and Capability for Assist-as-Needed Upper-Limb Robots uses a dual multimodal MHE architecture with an attribution gate to distinguish between changes in intended motion and changes in effective voluntary torque capability, successfully resolving ambiguity between the two update channels. This is complemented by SPEAR, which proposes a framework that transforms vision language models into efficient preference oracles by using query selection and adaptive reward transitions to improve success rates on contact-rich manipulation tasks.

Finally, there is the effort to consolidate skills in vision language action models. Multi-Teacher On-Policy Distillation for Flow-Based Vision-Language-Action Models proposes a two level distillation framework that uses teacher velocity supervision alongside student interaction data to consolidate domain skills into a single policy, achieving success rate gains on benchmarks like LIBERO and CALVIN.

The work by KwangBin Lee on Diffeomorphic Hidden-State Steering is particularly important because it addresses how small errors in imitation learning policies can compound during real-world execution, which directly impacts robustness. This paper proposes Diffeomorphic Hidden-State Steering to improve policy robustness by learning an invertible nonlinear coordinate map from expert hidden states. The method works by attenuating coordinates with low expert variance more strongly than those with high variance and then mapping these modified coordinates back to the hidden state space, which allows steering directions to vary with the current state while preserving all original coordinates; this approach leads to higher mean success under observation noise across four simulated manipulation tasks compared to unmodified Transformer policies.

The findings from Guoping Pan Xinmiao Yu Wannan Cao Jincheng Li Baolin Jiang Yizhe Li Yi Cheng Houde Liu's Trajectory-based Diffusion from One-shot Human Video for Generalizable Manipulation show how trajectories can be extracted from human videos using plug-and-play 6D pose estimation to create a modality agnostic intermediate representation. This framework enables generalizable manipulation by focusing on task relevant objects, achieving near one hundred percent success rates in simulation tasks and enabling cross embodiment generalization in real experiments; this contrasts with the work by Danqi Zhao Keyu Chen Wenchao Sun Minglei Huang Jianhe Zhang Yan Zhu Hao Cheng Zifan Cao's Plans Are Not Executions: Adapting Frozen World-Model Planners to Controller-Induced Execution Shift, which shows that directly substituting predicted execution trajectories yields only a small improvement over adaptation methods like ExecutionBridge.

Junlong Shen Shunsuke Shigaki's research on when a robot can trust reversing wind to find an odour source identifies distinct geometric conditions for successful wind guided search, linking heading correctness to net odour transport and intermittent contact; this is related to the challenges in maintaining tracking in dynamic environments, as seen in VP-EVT: Active Visibility Preservation for Embodied Visual Tracking, which uses language guided perception and visual history combined with ego motion to predict whether continued following will sustain an obstruction.

Jon Skerlj Sami Haddadin's work on Sub Millisecond Reflexes for Continuous Physical Human Robot Interaction demonstrates how a sub millisecond reflex engine can enable deterministic switching between control modalities, reducing execution time and peak trajectory jerk during multi stage tasks; this rapid adaptation capability is also relevant to TRACE: Typed Reference Aware Conversational Editing for Humanoid Motion Diffusion, which uses typed kinematic quantities to ground motion generation and allows for reference aware revision of movements based on language requests.

The most significant work here is MAIA because it tackles the core problem of anticipating interaction anomalies before they become obvious during contact rich manipulation. This is crucial because reactive monitoring only tells you something went wrong after the fact, whereas MAIA aims to predict future failures by learning nominal dynamics and predicting visual and force responses from planned actions.

This anticipation capability is built upon a real world multimodal dataset called CRIA, which consists of two thousand eight hundred fifty-two trajectories covering eighteen different contact rich manipulation tasks with frame level annotations for when anomalies start and evolve; this allows MAIA to train its lightweight anomaly head to predict impending issues, achieving better discrimination scores than existing baselines while providing earlier warnings.

The underlying mechanism relies on the vision action force world model that learns the expected interaction dynamics of a task, which is then used by MAIA to forecast what visual and force responses should occur from a planned action; this predictive modeling helps improve end to end task success rates significantly, boosting them from fifty two point six percent up to sixty one point five percent without needing to retrain the main manipulation policy.

The framework's ability to predict future consequences is also linked to how policies are trained in vision language action learning, where researchers have explored whether allocating training samples toward auxiliary objectives like latent action or future vision prediction rather than just action prediction can improve learning under a fixed budget; this idea is explored in the work <a href="/papers/2604.01730">MAIA</a>, which shows that modeling future physical consequences is valuable for proactive anticipation.

Another related area involves improving trajectory optimization for complex mechanical systems, as MAGMA introduces an Affine Geometric Heat Flow based optimizer that handles high dimensional problems with constraints better than current solvers by using a primal dual flow approach; this mathematical rigor complements the empirical learning approaches seen in <a href="/papers/2604.01730">MAIA</a>, showing different ways to tackle complexity in robotics.

Furthermore, efficiency in deploying these models is addressed by FastVLA, which introduces task aware and temporally adaptive visual token compression to reduce the computational load of vision language action models during closed loop deployment on robots. This compression technique can cut inference latency by up to thirty two percent on a physical robot while maintaining performance comparable to existing methods.

Finally, for tasks involving physical state changes, persistent profile guided reset free recovery shows that predicting multiple outcomes over repeated gait execution helps guide short online probes without resetting the system, leading to better final gait quality in damage recovery scenarios.

The work by Mohammad Dastranj Jouni Mattila AKMPC is particularly important because it tackles the fundamental problem of determining unique physical outputs from closed chain mechanisms where internal reactions can become indeterminate even with fully specified motion. This paper develops a modular inverse dynamics workflow that combines path based motion reconstruction with the recovery of uniquely determined actuator and joint wrench outputs; while this method uses analytical second order differentiation of assembled transformation paths to supply passive velocities and accelerations on a regular closure branch, it then assembles spatial Newton Euler balances over the physical body and joint graph to identify unique outputs using nullspace and dual criteria.

This is supported by the work of Cesar E Ramos Chuquiure Ryan Tsai Yutong Zhou Zihang Huang Vansh Thakur Akshunn Trivedi Zachary Lin Pearl Lin Rikai Yang Aallen Enninful Adu Ram Vasudevan LEAP, which addresses the difficulty of generating dynamically feasible motions for robots by solving a large nonlinear optimal control problem involving contact wrenches under friction cone constraints. LEAP introduces a primal dual gradient flow of an augmented Lagrangian to compute solutions, and it is noted that across various locomotion benchmarks on five robot platforms, it is the only method tested that solves every problem and achieves substantial speedups over other solvers.

Moving down in significance, Gorkem Gemalmaz Harun Tolasa Volkan Patoglu Perceptual Comparisons of Viscoelastic Renderings investigates whether users perceive fractional order viscoelastic models as more realistic than integer order models when rendering materials with memory effects. The study found that the fractional order standard linear solid model is perceived as significantly more realistic than both integer order alternatives under both parameter selection approaches; this contrasts with the work of Mani Amani Mohammed Alyaseen Zhirui Dai Ki Myung Brian Lee Nikolay Atanasov Jorge Cortes BubbleCBF, which constructs constructive safety constraints for second order systems by using quadratic geometry to build collision free spheres, proving that the resulting controller is well defined and forward invariant for differential drive robots.

Finally, Federico Cecati Yang Wu Florian Dorfler Marco Liserre Xiongfei Wang Elisabetta Tedeschi Stability Analysis in IBR Dominated Grids discusses the challenges of analyzing stability in converter dominated grids by examining model based and data driven methodologies; this paper explores how different modeling tools handle the interplay between static, small signal, and large signal phenomena, though it notes that challenges remain regarding data completeness and physical interpretability when using data driven approaches.

The work on safety filters for control barrier functions matters because it directly addresses how to ensure physical systems remain safe even when operating near their limits. AKMPC investigates the dynamical properties induced by these filters in a simple planar linear system, showing that even with a stable nominal controller and an affine constraint, the filtered system can still oscillate or diverge while maintaining safety; this means we need to precisely define the conditions under which stability is guaranteed for these systems.

The results from RV-C cubed BF show how visual safety certification can be achieved using only scale free relative motion encoded by monocular optical flow, directly encoding collision cone geometry without needing a full world model; this approach is significant because it bypasses the need to reconstruct complex obstacle positions for collision avoidance.

The PDR Sensor (PDRS) module described in PDRS simplifies localization by creating a plug and play black box unit that fuses inertial sensor data, achieving low latency and modest power consumption; this lowers the technical barrier for integrating navigation into wearable tech, contrasting with more complex algorithmic designs that often lead to long development timelines.

Moving toward manipulation, VaM introduces Value as Memory, a framework that uses value caching and online reinforcement learning to improve generalist robot policies by combining offline demonstrations with online updates, yielding significant success rate gains in long horizon tasks; this method shows how memory can accelerate policy adaptation in complex manipulation scenarios.

In terms of search efficiency, TOER proposes a new formulation for time optimal ergodic retrieval that models detection as a Poisson process, achieving up to nineteen percent shorter mission durations compared to existing ergodic baselines when the return phase is costly; this suggests a better way to plan search trajectories in environments where returning is expensive.

The most significant development is the proposal of a complete end-to end neuromorphic paradigm for micro aerial vehicle visual navigation in complex environments because it directly addresses the critical need for robust low power autonomy in energy constrained systems. This work shows that by integrating neuromorphic vision and spiking neural networks, this method can achieve navigation performance comparable to traditional artificial neural network approaches while drastically cutting power consumption.

This feasibility is supported by the fact that the spiking perceptual module uses only nineteen point five seven percent of the power of its ANN counterpart, and when deployed on Intel Loihi, it consumes only twelve point three seven percent of what an ANN running on NVIDIA Orin NX requires; this sets a new benchmark for energy efficiency in visual processing.

Further context comes from the work on location based safety management with wearable seamless indoor outdoor positioning, which shows that resource constrained wearable tags can achieve mean errors of one point three two meters indoors and three point three zero meters outdoors using LTE Cat M communication; this demonstrates practical application for real time safety monitoring in mixed environments.

In terms of control methods, MR-MPCC presents a Model Predictive Contouring Control framework designed for velocity controlled robotic manipulators to maintain small path tracking errors even when the reference frame is moving; this is crucial for tasks like machining and outperforms conventional time indexed tracking MPC by achieving lower geometric path following error under moving reference disturbances.

Another area of advancement involves optimizing energy allocation through speed trajectory optimization for sustainably powered mobile robots, where the study uses clarity as an information metric to derive optimal speed profiles that maximize information gain over time; this research illustrates how robots can intelligently manage limited energy resources for long duration missions.

Finally, the paper on de risking robotic welding for in space metal joining through physics grounded digital twins provides a model to screen welding conditions before physical tests by using Gaussian process surrogates to predict weld pool behavior under reduced gravity; this simulation approach helps mitigate risks associated with testing in space environments.

Today's papers

The papers

Important terms

KODA
This work converts human demonstrations into training data for whole-body vision language action by using a reusable canonical trajectory to guide robot view synthesis and generate specific controller supervision.
LIN-SLAM
A neural implicit backend for LiDAR inertial SLAM that achieves real-time, top accuracy online. It couples high-rate geometric estimation with neural implicit tracking for practical mobile robot navigation.
MAIA
A framework that anticipates interaction anomalies before they happen during contact manipulation by learning nominal dynamics and predicting visual and force responses, improving success rates significantly.
Fast-VLA
Introduces a mechanism where a frozen foundation model is used once per episode to establish fixed task memory. This cuts policy inference time substantially by allowing chunk prediction without repeated context recomputation.