The authors propose the Just Action Model (JAM) framework, which breaks down the differences between VLA and WAM into three separable choices: backbone initialization, the loss objective on future observations, and whether that loss mechanism is retained at inference. JAM unifies both into a structure of a pretrained visual backbone plus an action expert, trained with an action loss and a foresight loss.
The authors propose a simulation-driven framework centered on a multi-task coronary perception network, MCP-Net, for robotic guidewire navigation, integrating coronary vessel segmentation, vascular topology extraction, and roadmap generation. The method generates simulated fluoroscopy images and simulation-derived ground truth from three-dimensional coronary CT angiography, including simulated guidewires, stenotic lesions of varying severity, and diverse vascular geometries. MCP-Net jointly performs vessel segmentation, guidewire-tip localization, and stenosis localization from a single fluoroscopic frame.
This study proposes a reliability-aware framework that connects polarity-time localization, geometric recovery, event-CAD refinement, and trajectory optimization in sequence, using a three-branch representation that combines positive events, negative events, and recent-event timing. On the official SPADES challenge test set, the complete system achieves a combined pose score of 0.0474, about 5.4% lower than PASAN+IRIS. Ablations show that base localization with uniform PnP scores 0.0772, while the full cascade reduces it to 0.0514; under identical single-frame observations, quality weighting lowers the score from 0.0549 to 0.0514.
The paper analyzes loss of plasticity (LoP) in isotropic Gaussian policies within Proximal Policy Optimization (PPO), noting that the policy standard deviation σ significantly affects the extent of LoP, with theoretical and experimental support. The experiments include LoP benchmark comparisons and real-world policy deployment on a 7-DoF robotic arm; the authors also use these findings to explain why common LoP mitigation methods are effective.
To address the need for autonomous agents in shared environments to guarantee joint compatibility, such as safety and task-completion requirements, this work proposes a bi-level structure: a regulator publishes seed agents, and each developer independently tests its own agent for compatibility with every seed-compatible agent, a property called "future-proofness."
This paper studies the failure of world-model uncertainty after a deployment policy changes. It proposes weighted conformal calibration to transport marginal coverage to the target policy distribution, and it bounds the weight error using the log-loss regret of a domain classifier.
ThinkWorld is a vision-language-action (VLA) learning framework that represents each action thought as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens that captures the intermediate dynamics connecting the current world to a future state and serves as a common action representation across embodiments.
The research team proposed a training recipe for in-hand object pose reaching with the LEAP dexterous hand. It trains in simulation without human data and, after system identification, transfers zero-shot to the real robot, using either motion capture or a single wrist-mounted monocular depth camera. In a real-robot evaluation with 4 objects and 4,800 unseen targets over 13.3 hours, the state-based policy reached 95.2% of targets within 2 cm and 0.4 rad; adding a learned residual safety filter with adjustable conservativeness reduced drops by 75% while still reaching 91.5% of targets.
RESPLE2 proposes a recursive multi-sensor state estimation framework based on non-uniform B-splines for 6-DoF motion representation, using the knot-clamping boundary property to predict control points directly from inertial kinematics. Prediction and reset are performed within an iterated extended Kalman filter, and the update step fuses LiDAR, camera, IMU, and leg kinematics at their native timestamps; a separate lightweight global optimization module refines control points using LiDAR-visual loop-closure constraints.
EgoThumb is a fully open-source, vision-driven supernumerary robotic thumb that uses a visual imitation learning policy to map first-person images from a palm-facing camera directly to continuous joint positions, enabling autonomous grasping and holding assistance. Its compact dual tendon-driven finger relocates the actuation unit to the forearm to reduce distal weight and support precise active joint control. The authors state that the design generalizes across different users and environments, that both hardware and software are open-sourced, and that the paper has been accepted to IROS.
The research proposes a differentiable moving narrow-gap modeling method that gates centering, alignment, and smooth collision terms by signed distance to a time-varying gap plane, turning sparse traversal events into dense pathwise objectives, and integrates this into a visual-feature-to-command policy trained in a single stage from random initialization, without a teacher, planner, or distillation. In simulation, path gradients clearly outperform score-function gradients, reaching a 99.3% success rate after a few minutes of training on a single GPU. The policy transfers directly to a real quadrotor using only onboard vision and inertial measurements, successfully traversing the dynamic narrow gap in 8 of 10 real-world trials.
The authors propose Constraint-Referenced Strategy (CRS), which projects the end-effector twist commanded by a policy onto the object's joint axis, separating progress along the joint direction from off-constraint waste. On 74 simulated articulated objects, binary success rate only shows a lower transfer failure rate without attribution, whereas the two CRS channels give each a distinct signature.
The paper proposes Adaptive Residual Correction (ARC), which learns residual corrections via reinforcement learning on top of a frozen imitation policy, using sparse human corrections to decide when to activate residual control and in which direction of the residual action space to explore. On three precision manipulation tasks with a real robot, ARC raises average task success from 33.3% to 75.0%; the human corrections define 2-3 dimensional task-relevant subspaces of the seven-joint residual space, retaining 96.0-97.7% of the squared correction magnitude.
The authors propose an analytic event simulator that synthesizes event streams directly from scene geometry and ego-motion without image rendering, generating 64x64 event observations in just 17 microseconds on a single GPU, over 400 times faster than state-of-the-art rendering pipelines. Reinforcement learning training for an event-camera obstacle-avoidance policy based on this simulator converges in under 30 minutes, three orders of magnitude faster than training on rendered events, and transfers zero-shot to ray-traced event simulation in Isaac Sim.
Why it matters: This work uses analytic event simulation to bypass image rendering, cutting event-camera reinforcement learning training time to under 30 minutes, and completes real-world deployment on a 56-gram drone.
The StabBot robotic stabbing system uses a pendulum Kalman filter to track pipe joints in stereo images, approaches at matched speed, and confirms the final few millimeters by touch. In 200 randomized jobs, StabBot completed every stab, versus 74% for a conventional fixed guide; median contact speed fell from 18.4 cm/s to 2.3 cm/s, and median peak head-pipe force fell from 11.3 kN to 1.4 kN.
The study measures, on a simulated bimanual tool-use task, which failure causes each sensing channel lets a diagnoser identify: a tool slipping out of its socket can be read on the tool-hand side; which hand was late requires information from both arms, while a sensor on one hand can only read that hand's own delay and part of the other hand's delay; and when the tool hand is late, passive sensing cannot determine whether the push force was also insufficient.
The paper proposes a framework that relies solely on distributed channel measurements from communication infrastructure to sense passive entities outside a robot's field of view and perform risk-aware classification. The method uses hyperdimensional computing (HDC) to embed measurements into high-dimensional representations without explicit channel reconstruction. Numerical simulations based on commercial ray tracing and validation in a ROS smart factory scenario show accuracy of up to 89% under severe occlusion while maintaining low computational complexity.
World Dreamer v5-lite is a multi-robot world action model that predicts separate action sequences for each robot through a shared world state. The model builds the world state in workspace coordinates from calibrated features of a pretrained video encoder, and its spatial size does not change as views are added; a vision-language model provides scene context and one query per robot, and the queries read the shared state and exchange pose-aware information while their action streams remain separate.
Ego-CEVR proposes a data-native hierarchical visual representation learning framework that assigns distinct learning objectives to global and interaction representations. The global module learns object relations, task state, and temporal progress through future prediction on egocentric video; the interaction module interprets wrist observations in global context and learns local alignment and contact cues through robot action supervision. Two residual paths feed both representations into the action policy simultaneously, and the global video module can be pretrained independently on large-scale human egocentric videos and videos from multiple robot embodiments, without robot action labels or paired wrist observations.
A team at Delft University of Technology in the Netherlands proposed a fixed-wing interceptor controlled by a neural network that maps filtered monocular vision observations directly to actuator commands; the policy was trained in a custom Python environment using reinforcement learning and proximal policy optimization (PPO). In simulation, the policy outperformed two classical guidance controllers under most conditions, and it has successfully intercepted maneuvering fixed-wing targets in real flight. This work provides validation for the autonomous control of fixed-wing interception drones.
Researchers propose an electric-field induction motor whose rotor is made of discrete circuit elements between a finite number of conductive vanes, overcoming the limits of a single continuous material in resistivity, permittivity, and circuit characteristics. This macroscale motor uses circuitry on the rotor surface to emulate the "deep bar" rotor design of magnetic induction motors, providing higher torque over a wide speed range. The study completed the analysis, design, and demonstration of the motor.
The study proposes MARL-Exo, a multi-agent reinforcement learning framework in which a full-body musculoskeletal model and a bilateral hip exoskeleton policy are co-trained in simulation in two stages; the exoskeleton observes only delayed hip kinematics and past commands. In simulation, the metabolic policy reduced modeled muscle metabolic power by 13.8% compared with walking without the device, and by 8.8% after adding mechanical-power shaping.
AuRaSim is a ray-tracing-based automotive FMCW radar simulation framework that generates coherent multi-channel ADC data for dynamic scenarios. It builds dynamic scenes with Blender, performs GPU-accelerated electromagnetic ray tracing with Sionna RT, and adds path-level complex-coefficient propagation, motion-aware Doppler modeling, and waveform-configuration-driven ADC cube synthesis. In a digital twin scenario, the simulated and measured range-Doppler maps achieve an RMSE of 0.147 and a PSNR of 16.93 dB.
The authors propose Quality-Controllable Policy (QCP), which enables inference-time control over individual quality attributes such as execution speed and motion smoothness, as well as their combinations. To extend the controllable quality range beyond the initial demonstrations, the authors design an iterative hindsight relabeling loop: collect online rollouts under target quality labels, add successful trajectories to the dataset, then relabel the accumulated trajectories in hindsight according to their measured quality and fine-tune the policy.
PDRVLA proposes a lightweight Physical Dynamics Module that extracts physical tokens from vision-language features to condition action generation, without online video generation. The module learns through direct prediction of future RGB, depth, and optical-flow features, with auxiliary supervision from a frozen pretrained video world model. All training data comes from public data and was expanded through human-to-robot synthesis, subtask annotation, and embodied chain-of-thought (ECoT) annotation.
Researchers propose VLA-Stem, a lightweight reinforcement-learning-based safety mechanism that does not fine-tune the underlying VLA policy; instead, it converts its action chunks into motion references and, at inference time, adaptively corrects execution using proprioception and local obstacle information. Across six constrained simulation settings, VLA-Stem achieved an average collision-free task success rate of 52.8%, versus 23.7% for the strongest safety baseline, while preserving the original policy's performance in obstacle-free environments; real-robot experiments also showed reduced collisions under static and dynamic obstacles.
The authors propose a fully automatic, target-free framework that jointly estimates camera intrinsics, radial-tangential distortion, and camera-LiDAR extrinsics from raw images and LiDAR scans. The method extends deep pixel-point correspondence-based calibration from rectified images to raw cameras through intrinsic- and distortion-aware correspondence learning, and iteratively couples correspondence prediction with nonlinear optimization during calibration. Zero-shot evaluations on an unseen KITTI configuration and on three real robotic platforms show that joint calibration improves extrinsic accuracy while recovering accurate intrinsics.
The paper proposes the Adaptive Failure-Informed Learning (AFIL) framework, which uses a pretrained VLA to generate failure trajectories online as adaptive negative guidance for diffusion- and flow-based VLA policies, avoiding the need for handcrafted failure modes or human intervention for recovery. The method jointly trains two action generators (DAGs) for success and failure on a shared vision-language backbone, and at sampling time adjusts the guidance strength according to the distance between the success and failure distributions at each diffusion step. On in-domain and out-of-domain, short-horizon and long-horizon robotic manipulation tasks, AFIL improves task success rate and robustness compared with existing VLA baselines.
This paper investigates event-driven spiking controllers implemented as spiking neural networks for trajectory tracking in underactuated robots, with the goal of stabilizing a prescribed periodic orbit. The authors reduce orbital stabilization to stabilizing the trivial solution of a periodic linear time-varying system via transverse linearization, and extend a Lyapunov-based greedy design originally developed for LTI systems to this periodic LTV transverse dynamics, thereby constructing a feedback controller. The method's effectiveness is demonstrated on the trajectory stabilization problem for the butterfly robot.
KEPT proposes a navigation method for enclosed spaces such as tunnels and mines, using conformal prediction to calibrate keepouts around mapped walls under degraded sensing, and adjusting the margin according to a label-free degradation severity estimate to maintain coverage. The planning and control layer enforces a braking-distance condition on the keepout at plan time and on every command, using the measured tracking residual as its only margin and never altering the keepout.
The authors propose Retrieve to Act (R2A), a test-time scaling framework that organizes successful demonstrations into an offline spatiotemporal motion primitive graph. At inference, a query-dependent graph neural network locates the current execution state and retrieves structurally compatible future motion primitives, which are then fused with the frozen VLA's predicted actions through dimension-wise gating.
The authors propose Local Spatial Matrix Memory (LSMM) based on Hadamard memory, which lets reinforcement learning agents store task-relevant information in position-based memory cells. On an autonomous maze exploration task, LSMM's maze coverage is comparable to strong recurrent networks and finite-history baselines, and it is the only architecture that remembers the maze layout, reducing reconstruction error to one third. The work targets the problem that humanoid and legged robots in complex environments need to remember information long after it leaves their field of view before it is used.
For a planar monoped robot under lunar gravity (1.625 m/s²), this study proposes a bi-level co-design framework: a differential evolution outer loop searches link lengths, and each candidate is evaluated for vertical jump performance by either a QP task-space controller (QP-TSC) or direct-collocation trajectory optimization.
The paper proposes a denoised counterfactual protocol that uses bitwise-reproducible branching, repeated continuations, held-out oracles, and common random numbers to measure the selectable space a world action model (WAM) can achieve when selecting candidate actions at test time, and builds a fail-gated selector on that basis.
This paper investigates two passive geometric approaches to reduce the acoustic signature of a flapping membrane wing and validates them on an exemplary wing. After optimizing the venation layout based on a dynamic snap-through model of the membrane wing, experiments measured a reduction of up to 8 dB in the stroke-reversal snap noise; an additive trailing-edge fringe inspired by hummingbird feathers further reduced the broadband swoosh by 3–4 dB without loss of lift. The paper notes that engineered wings are 15–30 dB louder than the animals they imitate.
Researchers trained four imitation-learning policies—ACT, Diffusion Policy, SmolVLA, and π0.5—on a low-cost SO-101 robotic arm using 12 to 200 teleoperated demonstrations, evaluating 1,340 episodes in total.
RAVEN is an active SLAM framework for quadrotors that unifies planning of exploration and active loop closure under localization uncertainty, and uses a priority parameter from 0 to 1 to adjust the emphasis between exploration efficiency and localization accuracy. The framework first identifies frontiers and active loop closure candidates and generates global tour plans, then evaluates them with a utility vector, and finally a perception-aware navigation planner generates smooth trajectories based on an online Fisher information field to reduce visual state estimation error.
The authors propose GAWM (Geometry-Aware World Model), a compact multi-task world-action model that predicts future observations in the latent space of a frozen DINOv3 encoder, with fewer than 100 million parameters and no robot or video pretraining.
This study proposes a state reconstruction pipeline for a hybrid rigid-continuum space manipulator whose distal link is a tendon-driven continuum robot. The method formulates the estimation problem as maximum a posteriori inference conditioned on joint encoders and vision measurements, transcribes it as a factor graph, fuses heterogeneous measurements with the rigid-soft kinematic model, and uses the Geometric Variable Strain (GVS) approach to obtain a minimum-order parameterization of the configuration. Closed-loop trajectory simulations verified the state reconstruction accuracy and computational performance.
This paper proposes a hybrid discrete-continuous multi-objective bilevel framework for the structure-actuation co-design of all-electric heavy-duty robotic manipulators: the upper level optimizes geometry and actuation topology, while the lower level searches for feasible configurations of motors, gearboxes, and screws for each joint. A system-level coordinator selects actuator packages based on hardware mass, co-design cost, and loss-aware supply energy, and feeds the resulting actuator masses back into the robot model until mass consistency is reached.