Daily Summary for 2026-09-25

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Dev: Welcome to the show!

Rosa: Today we have a special show for you.

The summary: Rosa: Welcome everyone to our review of September twenty fifth, twenty twenty six research. Today we focus on improving robot movement when the shape is unknown.

Dev: MorphIK uses the robot's shape to condition neural inverse kinematics for unknown robots by figuring out the structure from what it sees first.

Taro: World Action Agent uses large vision-language models to rehearse actions in a simulated world before performing them, improving task handling through experience.

Rosa: RAPID focuses on agentic programming directly from demonstrations, learning how to program robots just by watching someone do the task.

Dev: Rolling-WAM deals with world action models incorporating rolling imagination to explore potential future actions dynamically during operation.

Taro: This feeds into Coding Agents for Generalized Task and Motion Planning Problems which aim to solve general planning issues.

Rosa: RAPID contrasts with uncertainty-gated exploration noise suppression in online reinforcement learning for flow-matching policies, tackling task collapse.

Dev: RotVLA is significant; it tackles controlling vision language action models by introducing a rotational latent action for better physical execution.

Taro: Learning to Navigate with Minimal Parameters decomposes visual navigation into closed-form geometric interfaces, reducing parameter count needed.

Rosa: Representation World Model learns states, transitions, and executable plans within a framework to allow agents to reason about their environment.

Dev: RAPTOR uses physics-informed solvers as a random-projection transient solver to solve physical problems with constraints from known laws of physics.

Taro: GridSFM presents a foundation model for solving AC optimal power flow problems, providing structured problem-solving for electrical engineering.

Rosa: Physics Guided Residual Reinforcement Learning for Humanoid Narrow Path Traversal uses physics to guide RL policies for safe movement in tight corridors.

Dev: This guided learning shows more stable traversal than standard methods, building on EgoSpeedUp's idea of mimicking human manipulation tempo.

Taro: BeyondRetarget learns executable humanoid motions directly from monocular video inputs, bypassing traditional modeling steps entirely.

Rosa: This complements novel view synthesis like M3GD, which uses multi-modal data to generate new geometric views for cameras and LiDAR systems.

Dev: So we have shape inference, action rehearsal, demonstration learning, and physics guidance across the board today.

Taro: It seems the focus is heavily on grounding complex models in physical constraints or structured representations.

Rosa: Precisely. The combination of these techniques is key to making these systems reliable in uncertain environments.

Dev: Indeed. The interplay between learned models and explicit physical guidance defines the cutting edge now.

Taro: A rich day for research covering perception, planning, and direct action learning across various domains.

Rosa: It certainly shows how diverse approaches can converge toward robust autonomous capabilities on the twenty fifth of September, twenty twenty six.

Dev: Let's move on to the next part of our review then. This was a busy session indeed.

Taro: Agreed. I look forward to discussing the next set of findings with you all soon.

Rosa: Until then, thank you for joining us for this research deep dive into robot intelligence.

Dev: Goodbye everyone and have a productive rest of your day.

Taro: Farewell and stay curious about the latest developments in robotics research.

Rosa: That's all for part one of our review today. We'll be back soon with more insights into this fascinating field.

Dev: Stay tuned for the next episode where we delve deeper into these concepts.

Taro: Until next time, keep exploring the possibilities in robotics research.

Rosa: Thank you for listening to this segment of our research review. This concludes part one.

Rosa: The Trajectory Induced Self Calibration work is key for locating targets when the robot's pose is unknown.

Dev: That uses the trajectory itself to calibrate, which connects conceptually with Free-Init for Doppler LiDAR systems.

Taro: Did you see Synthetic Enclosed Echoes? It creates a dataset bridging simulated and real sonar data.

Rosa: Yes, it helps train Self Adaptive VLA in more realistic scenarios. This shows a trend toward robustness.

Dev: StageCraft addressed failures from distractions in virtual labs for visual learning agents by improving execution awareness.

Taro: That relies on the coordinate-independent robot model identification first to feed into StageCraft.

Rosa: GenPHRI is also important, exploring agentic generative simulation for physical human-robot interaction.

Dev: We also made progress on sampling-based MPC for Double-Pendulum Sway Suppression on a shipboard crane using MuJoCo.

Taro: FingerViP focuses on learning dexterous manipulation by incorporating fingertip visual perception for contact understanding.

Rosa: MPC-Injection is significant because it biases off-policy RL toward behaviors aligning with what a controller would induce.

Dev: That builds on memory-guided agents steering latent agents into reliable manipulation primitives.

Taro: Modeling robot velocity fields as probability fields helps make motion planning more robust, connecting to ContactWorld.

Rosa: And for cloth manipulation, we are using inference-time simulator-in-the-loop refinement to fix model inaccuracies.

Dev: Human-in-the-loop geospatial annotation speeds up training data construction for field deployed UAV systems.

Taro: OCC4M aims to give spacecraft long-horizon manipulation by incorporating object-centric four dimensional memory.

Rosa: So, we have work on localization, simulation data, execution awareness, and advanced manipulation skills.

Dev: It looks like the overarching theme is making robotic systems more robust through better planning and handling uncertainty.

Taro: Exactly. The focus is on improving motion planning, learning from demonstrations, or handling sensor uncertainty.

Rosa: Right. And we are also looking at how to bridge learned policies with physically executable control strategies via MPC-Injection.

Dev: That seems like a major step forward for practical deployment of these complex systems.

Taro: It is certainly pushing the boundaries of what we can expect from autonomous operation in these environments.

Rosa: Indeed. The progress across all these areas shows a clear direction for more capable robots overall.

Dev: I think the next phase will involve scaling up the successful integration of these individual techniques together.

Taro: That sounds like a logical next step for synthesizing this diverse research pipeline into a unified system.

Rosa: Agreed. We have a lot of important, concrete work to synthesize from this day's findings.

Dev: Let's keep tracking how these components interact in the coming weeks.

Taro: I look forward to seeing those interactions materialize in future experiments.

Rosa: Definitely. This is a very productive review session.

Dev: It certainly keeps us engaged with the cutting edge of robotics research today.

Taro: It does, and it shows how interconnected these different research threads truly are.

Rosa: Precisely, the connection between motion planning and sensor uncertainty is becoming clearer.

Dev: It’s a complex landscape, but the solutions we are finding are increasingly tangible.

Taro: Tangible progress in handling real-world execution issues is what really stands out this week.

Rosa: I agree. StageCraft seems to be addressing that execution awareness gap effectively.

Dev: And GenPHRI opens up exciting possibilities for safe, intuitive human collaboration too.

Taro: It seems like we are making steady, verifiable progress in several critical areas simultaneously.

Rosa: Yes, and the data collection methods are also improving rapidly through human involvement in annotation.

Dev: So we have a solid foundation for robustness across perception, control, and learning mechanisms.

Taro: A very comprehensive picture of today's significant contributions to the field.

Rosa: So, we have work on tendon-driven continuum robots with modular stiffness and self-pose estimation.

Dev: That builds on complex physical interaction by controlling stiffness and estimating position without external sensors.

Taro: And we also have OA-MPPI for UAV flight, handling occlusions during navigation.

Rosa: That connects to the work on Excitation-Supervised Self-Calibration for range-bearing relays under uncertainty.

Dev: Right, and SCoCaT addresses spacecraft docking using success conditioned reinforcement learning.

Taro: The most significant morning work was streaming deep reinforcement learning for adaptive continual learning in robotics.

Rosa: That directly tackles robots needing to learn new tasks under communication constraints while operating.

Dev: It builds on Streaming-WAM, which developed an action-conditioned world-action model for asynchronous manipulation.

Taro: Then there's Koopman-accelerated model-based diffusion for real-time control to speed up action planning.

Rosa: FlyCNS focuses on connectome-grounded information organization for communication-constrained embodied control.

Dev: TactileStep looked at sole tactile learning to regulate foot interaction on uneven surfaces in humanoid locomotion.

Taro: RoboRecover benchmarks robot policy recovery under execution deviations, connecting to ActGaze's action-grounded gaze learning.

Rosa: The online adaptation of simulation models via closed-loop systems is very crucial for real-world reliability.

Dev: That involved testing alignment techniques for lifting oversized objects and outcome-sensitive motion search for impact catching.

Taro: We also have a simpler torque observation alignment method for zero shot sim to real grasping with direct drive grippers.

Rosa: Today's lucky papers are: MorphIK Morphological Conditioned Neural Inverse Kinematics for Unknown Robots.

Dev: World Action Agent Harnessing VLMs for Robot Manipulation via World Action Rehearsal.

Taro: Underwater C3-JEPA An Object-Centric Cross-View World Model for ROV Salvage.

Rosa: Coding Agents for Generalized Task and Motion Planning Problems.

Dev: Rolling-WAM World Action Models with Rolling Imagination.

Taro: RAPID Robot Agentic Programming from Demonstrations.

Rosa: Uncertainty-Gated Exploration Noise Suppresses Task Collapse in Online RL Fine-Tuning of a Flow-Matching Vision-Language-Action Policy.

Dev: Continuous Online Fault Detection for Mobile Robots via Adaptive Edge Models.

Taro: Decoupled Early Exits for Task-Dependent Compute Allocation in Flow-Matching VLAs.

Rosa: RAPTOR RAndom-projection Physics-informed Transient sOlveR.

Dev: GridSFM A Foundation Model for Solving AC Optimal Power Flow.

Taro: Free the Language Model From the Vision Encoder: Semantic Serialization as a Perception Interface for Small Language Models.

Rosa: RotVLA Rotational Latent Action for Vision-Language-Action Model.

Dev: Learning to Navigate with Minimal Parameters: Decomposing Visual Navigation Through Closed-Form Geometric Interfaces.

Taro: Representation World Model Learning States, Transition and Executable Plans in Representation.

Rosa: FMCW-LIO A Doppler LiDAR-Inertial Odometry.

Dev: Free-Init Scan-Free, Motion-Free, and Correspondence-Free Initialization for Doppler LiDAR-Inertial Systems.

Taro: EgoSpeedUp Transferring Human Manipulation Tempo to Robot Policies.

Rosa: BeyondRetarget Learning Executable Humanoid Motions Directly from Monocular Video.

Dev: M3GD Multi-Modal Multi-View Geometric Diffusion for Camera--LiDAR Novel View Synthesis.

Taro: Self-Adaptive VLA for Robust Robot Deployment.

Rosa: Synthetic Enclosed Echoes A New Dataset to Mitigate the Gap Between Simulated and Real-World Sonar Data.

Dev: Trajectory-Induced Self-Calibration for Hidden-Target Localization Through an Unknown-Pose Range-Bearing Relay.

Taro: Physics-Guided Residual Reinforcement Learning for Humanoid Narrow-Path Traversal.

Rosa: Object-Reconstruction-Aware Whole-body Control of Mobile Manipulators.

Dev: Coordinate-Independent Robot Model Identification.

Taro: Sampling-Based MuJoCo MPC for Double-Pendulum Sway Suppression on a Shipboard Crane.

Rosa: StageCraft Execution Aware Mitigation of Distractor and Obstruction Failures in VLA Models.

Dev: GenPHRI Agentic Generative Simulation for Physical Human-Robot Interaction.

Taro: RHINO-AR An Augmented Reality Exhibit for Teaching Mobile Robotics Concepts in Museums.

Rosa: FingerViP Learning Real-World Dexterous Manipulation with Fingertip Visual Perception.

Dev: Large-Scale Continuous Occupancy Mapping via Variance-Weighted Submap Joining.

Taro: GUIDE Goal-Initialized Directional Understanding for End-to-End Legged Navigation.

Rosa: ContactWorld What Representations Matter for Vision-Tactile Latent World Models in Contact-Rich Manipulation.

Dev: Flow as Flow Modeling Robot Velocity Fields as Probability Velocity Fields for Flow-Based Object Manipulation.

Taro: Enabling Robust Cloth Manipulation via Inference-Time Simulator-in-the-Loop Refinement.

Rosa: MPC-Injection Biasing Off-Policy Locomotion RL Toward Controller-Induced Behavior Basins.

Dev: Harness VLA Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents.

Taro: Implicit Behavior Coordination from Sub-Task Demonstrations by Exploiting Overlap-Induced Multimodality.

Rosa: Know Your Body A Harness for Direct and Self-Improving Robot Control with VLMs.

Dev: Morphometric Imitation From Morphology and Contact Aware Hand Retargeting to Sim-to-Real Visuomotor Policy.

Taro: OA-MPPI Occlusion-Aware Model Predictive Path Integral Control for UAV Flight.

Rosa: Excitation-Supervised Closed-Loop Self-Calibration and Target Seeking for an Unknown-Pose Range-Bearing Relay.

Dev: SCoCaT Success Conditioned Constrained Reinforcement Learning for Spacecraft Docking.

Taro: Tendon-Driven Continuum Robot with Modular Stiffness and In Situ Self Pose Estimation.

Rosa: TAPESIM Efficient Simulation of Adhesive Tape Dispensing for Robotic Manipulation.

Dev: Human-in-the-Loop Geospatial Annotation for Rapid Dataset Construction in Field-Deployed UAV Systems.

Taro: OCC4M Object-Centric 4D Memory for Spatiotemporal Reasoning in Long-Horizon Manipulation.

Rosa: An Analysis of Streaming Deep Reinforcement Learning for Adaptive Continual Learning in Robotics.

Dev: FlyCNS Connectome-Grounded Information Organization for Communication-Constrained Embodied Control.

Taro: Koopman-Accelerated Model-Based Diffusion for Real-Time Robot Control.

Rosa: Streaming-WAM Action-Conditioned World-Action Model for Asynchronous Robot Manipulation.

Dev: A Field-Deployable GNSS-based Navigation Stack for Outdoor Mobile Robots.

Taro: RoboRecover Benchmarking Robot Policy Recovery under Execution Deviations.

Rosa: ActGaze Learning Action-Grounded Gaze through Counterfactual Visual Interventions for High-Precision Manipulation.

Dev: TactileStep Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion.

Taro: Online Sim-to-Real Adaptation via Closed-Loop System Modeling.

Rosa: Fly Drive Reconfigure A Modular Reconfigurable Aerial-Ground Platform for Field Operations.

Dev: CALM Current Aligned Link Manipulation for Single Arm Oversized Object Lifting.

Taro: Outcome-Sensitive Motion Search for Impact-Aware Dexterous Catching.

Rosa: That concludes our review for today. Join us next time when we cover: MorphIK, World Action Agent, and Underwater C3-JEPA. Good day to you all. Goodbye.