The arXiv paper Toward Recursive Self-Improvement in Physical AI (arXiv:2610.29677, 93 pages, 15 figures, 21 tables) positions recursive self-improvement (RSI) within the robot learning framework and traces pathways from improving robot capabilities to improving the processes that generate them.
The paper connects research on an agent's acceptance of and avoidance of gaze with the affective viscosity model, which describes how negative emotions accumulate, persist, are suppressed, and are expressed over time. It argues that avoidance should not be seen merely as an interaction failure, but as a relational signal marking the boundary between humans and agents. On this basis, the paper proposes a design perspective in which agents maintain, suppress, or express negative states depending on social relationships and interaction histories. It discusses design cases including triadic jealousy, discomfort suppression, patience timing, and spatiotemporal stickiness, and then raises questions about whether interventions can be traceable, limited, contestable, and repairable.
Reading the Room (RR) proposes a global-local method for registering scale-ambiguous sparse semantic VSLAM maps to floor plans of partially built environments. Its core is a single robust, differentiable, class-balanced semantic-geometric energy function that runs through the entire pipeline. Evaluated on 33 sequences from 3 construction sites, it improves mean absolute trajectory error by 4.5 times over the strongest baseline in the general unconstrained setting, rising to 7.6 times when anchor correspondences are provided.
The arXiv paper Watch Your Step proposes a lightweight foothold prediction model (FPM) that predicts future footholds from proprioception and velocity commands alone, and samples the elevation map only around those footholds, letting perception anticipate foot placement ahead of the policy.
The study proposes a multistage reinforcement learning method that produces a single end-to-end control policy for commanded point-to-point jumping by a quadruped robot under lunar gravity. The policy architecture is simple and relies only on mission-compatible onboard sensing; in simulation, landing errors are smaller than the robot's body size. The work also analyzes torque and power demands and characterizes performance degradation under extreme outlier events, including control-update freezes of up to 1 second during flight. In hardware transfer tests, the policy successfully completed a vertical jump under lunar-gravity offloading, and measured torque and power profiles matched the simulation and stayed within hardware limits.
The study reveals the kinematic conditions under which Jacobian-based performance indices of serial robotic manipulators, such as manipulability and stiffness, are independent of the first and last joint angles, i.e., joint-angle invariance (JAI). This property systematically reduces the dimensionality of the effective configuration space and simplifies performance evaluation. In two scenarios, workspace performance map construction and task layout optimization, the JAI strategy reduces computational effort and storage requirements without compromising effectiveness. The work has been accepted to IROS 2026.
The paper proposes ContactFlowFilter for estimating the number, locations, and force vectors of active contacts on a robotic arm equipped only with joint torque sensors and a base force/torque sensor. The method uses a flow matching model as an auxiliary proposal distribution for a particle filter, leveraging the particle filter's recursive structure to capture temporal dependencies while tracking multimodal beliefs and detecting contact addition and removal events.
An actuator-constraint-aware master-slave teleoperation method that jointly models master-side gravity compensation, master-slave joint mapping, and slave-side differential wrist control, addressing the problem that independent motor saturation in differential wrists disrupts pitch-roll coordination and that fixed mappings struggle with abrupt tracking changes caused by gravitational loads and actuator constraints.
The paper proposes SERA (Self-Evolving Robotic Agent), which organizes a high-level VLM as a robotic agent with an editable harness, paired with a fixed VLA executor for low-level control, to handle non-Markovian dependencies in long-horizon manipulation.
The paper proposes the first computational approach for designing LEGO walkers that support multiple gaits beyond stable walking. The method has two parts: first, a specially designed gear-linkage mechanism controls the foot trajectory of each leg, incorporating a pair of 3D-printed non-circular gears working with a LEGO Technic mechanism; second, each leg is driven by an independent motor, with a simple yet flexible control algorithm synchronizing the timing of multiple leg motions.
S-Rolling is a steerable rolling gait for twistable snake robots that combines low-sweep-height locomotion with planar maneuverability. Based on an S-shaped planar backbone curve, its head and tail segments roll at independently controlled angular velocities to produce different motion modes.
The study proposes a linear twisting function that distributes torsional load across the entire body of a snake robot, replacing the existing rolling vector method's approach of concentrating it in two middle twisting joints. Under matched turning speed and total output power, the new gait reduces the peak output power of a single twisting joint by 42%. Real-robot experiments show it can move on various terrains that the existing rolling vector method cannot traverse, and is up to 49% faster than the traditional sidewinding gait without twisting.
The paper proposes A2I-Calib++, a kinematics-driven, noise-robust active calibration framework for multi-IMU systems mounted on legged robot bodies. It can estimate temporal offsets, extrinsic rotations and translations, and inertial biases without requiring artificial calibration targets or complex parameter tuning.
This framework combines flow-matching-based path planning with model predictive path-following control for mobile robot navigation. The flow matching model captures multimodal path distributions, while the predictive controller satisfies state and input constraints online and avoids obstacles, and event-triggered replanning feeds control solutions back into subsequent path generation. Simulation results show that the framework performs robustly in cluttered environments.
The paper proposes PAL (Plan first, Act Later), a hierarchical memory framework for vision-language robot manipulation. A high-level vision-language model (VLM) converts streaming visual observations into a subtask_list, and at execution time selects the currently relevant subtask from it to pass to a low-level vision-language-action (VLA) policy.
The paper proposes Relational Geometry Forcing (RGF), which injects geometric priors from a frozen 3D foundation model into World Action Models (WAMs) by aligning pairwise feature relations instead of pointwise feature alignment, avoiding the degradation of action learning caused by dense geometric targets over-constraining compact action features.
The authors propose TIDE (Training In DEployment), a human-in-the-loop online post-training system that uses a provenance-conditioned supervised learning framework to update the policy online from human interventions and policy rollouts during deployment, without learning a critic, reward model, or action scorer.
The authors propose DepState-JEPA, a dual-stream RGB-D joint-embedding predictive architecture that connects spatial understanding with action-conditioned prediction through a shared latent state, DepState. Built on V-JEPA 2.1, the framework uses depth both as a complementary observation stream and as the grounding for metric 3D space, and constrains state learning with ray-conditioned readouts, object-trajectory supervision, and calibrated forward kinematics.
NavBind proposes a method for diagnosing whether navigation vision-language-action (VLA) models truly follow instructions: with the navigation context fixed, it checks whether changing the instruction affects the probability distribution of actions. The authors recombined route-aligned instruction segments from R2R-CE and RxR-CE to construct 3,252 route-fork contexts, each pairing two full instructions that share a description of the route already traveled but specify different branches, and evaluated them under the same agent state and navigation history.
SMR is a streaming neural motion retargeter learned from semantic-conditioned physics experts. It converts human motion into robot task-space targets, which the physics experts track in simulation, and then uses a strictly causal Transformer to learn their executed trajectories. Each new frame uses only current and past human observations to generate a robot motion reference for a downstream whole-body controller. Experiments show that the semantic-conditioned physics experts outperform direct-optimization retargeting in semantic fidelity and physical validity, and that the resulting streaming retargeter performs better in closed loop under a unified whole-body controller than kinematic and physics-refinement baselines.
This study proposes that the generalization performance of world action models (WAMs) diverges significantly depending on the representational geometry of their generative backbones, and quantifies this prior using nuisance orthogonality (the degree to which features are orthogonal to the in-distribution manifold under out-of-distribution visual variations). Experiments across various image and video backbones find that higher orthogonality before training is associated with stronger robustness to visual variations after WAM training, enabling prospective prediction of policy generalization before training. Analysis of robust WAMs also reveals suppressed feature displacement and robust action generation.
SimPT is an automated simulation framework that integrates policy rollouts, corrective data collection, policy training, and standalone evaluation for human-intervention-based vision-language-action (VLA) post-training.
An arXiv paper proposes a lightweight state-dependent objective coordination method for end-to-end balance and navigation of two-wheeled inverted-pendulum robots. The method introduces a pitch-angle-conditioned scalar coefficient β(s)∈[0,1] that dynamically adjusts the relative weight of balance stabilization and navigation objectives during training, without requiring an additional control hierarchy or attitude-stabilization controllers such as PID or LQR.
ACT-Flow is an online learning framework that combines an adaptive Action Chunking Transformer (ACT) with a conditional residual flow policy, allowing ACT to absorb successfully executed actions through behavioral cloning, with residual targets recomputed relative to the updated prior.
PHD-Hand is a linkage-driven humanoid dexterous hand with three drive degrees of actuation (DOA) per finger. It uses a unique flexure hinge and elastic drive rods made of spring steel to simplify the transmission, avoiding the need for many miniature moving parts, and can be manufactured with budget-level 3D printers, lowering the barrier to use and development of dexterous hands and bringing its size closer to that of a human hand. A prototype has been used for grasping and object manipulation evaluation, and the paper also presents the latest research results on the hand's kinematics and control. It has been accepted by IEEE-RAS Humanoids 2026.
SLoSa is a neuro-symbolic compression and fine-tuning framework that builds a safety-aware dataset using classical planners, a teacher large language model, and formal verifiers, then applies supervised fine-tuning and reinforcement learning fine-tuning guided by a verification-based reward machine to train small, low-power, safety-aware robotic task planning models.
The authors' team has released GGraph, an open-source GPU-accelerated graph library for robot deployment and large-scale learning. It supports real robots through ROS 2 interfaces and supports vectorized GPU policy training, with built-in support for tasks including 3D navigation, exploration, and loop-closure integration. In single-robot operation, graph updates take milliseconds; in large-scale training, it supports hundreds of parallel robots training graph-based exploration policies, reducing wall-clock training time by an order of magnitude.
HyperDrone is a quadrotor simulator built for high-performance batch rendering, over 20 times faster than existing state-of-the-art batch quadrotor simulators.
COWBOY is a scalable sim-to-real framework for learning contextual whole-body manipulation. On whole-body dexterous grasping tasks, it transfers zero-shot to hundreds of objects and diverse in-the-wild scenes, including cluttered tabletops, shelves, drawers, and other unstructured environments.
A study compared four preference aggregation methods—reward-level, advantage-level, loss-level, and PCGrad—across two humanoid robot control frameworks, finding that no aggregation level is optimal across all environments and metrics. The experiments used Preference Reward Responsiveness (PRR) combined with coverage and utility metrics; results show that early aggregation produces a wider range of behavioral responses but introduces more unintended objective interference, while later aggregation improves control precision at the cost of some responsiveness.
The study proposes the Context-Aware Navigation Explanation framework, which uses a large language model combined with perceptual context and planned trajectories to generate natural-language explanations for assistive robot obstacle avoidance and navigation. The system is deployed on a Unitree Go2 platform in indoor navigation scenarios to evaluate how different explanation context configurations affect user trust and explanation satisfaction. Results show a strong positive correlation between explanation satisfaction and trust, and that the configuration combining obstacle information with future trajectory produced the best explanations.
This study proposes a general learning framework for nonlinear, nonconvex dynamic games, applied to multi-agent cooperative iterative tasks with shared constraints. It uses previously executed feasible trajectories to construct a sampled joint safe set and player-specific cost-to-go functions.
The paper proposes the Flexible Feature Editing Pipeline (FLEX), which uses a language model to iteratively correct a robot's internal representation of task features based on natural language corrections through three complementary operations: Add, Generalize, and Mask.
The paper proposes the inr2action framework, which fits a neural implicit representation to each robot manipulation episode starting from a shared meta-learned initialization; the resulting weight delta serves as memory, with a size that does not vary with episode length and stays fixed from one hundred frames to one hundred thousand frames.
Real2Sem2Real (R2Sem2R) proposes a real-semantic-real pipeline that turns a single monocular RGB human demonstration video into thousands of robot demonstrations covering multiple semantic object instances within the same category. The method tracks object SE(3) poses using DINOv3-distilled Gaussian splats, outperforming existing methods and FoundationPose with estimated metric depth, and maps the human hand to a feasible robot grasp that can be transferred to digital twins.
T-WLA is a World-Language-Action architecture that uses shared visual memory to guide both action generation and tactile correction, addressing the problem in contact-rich manipulation where contact conditions change before the visual policy updates.
VeriCalib formulates target-based LiDAR-camera extrinsic calibration as a verifiable estimation problem and outputs a calibration state indicating whether to accept the result or collect more data. It uses a structured covariance to decouple rotation and translation accuracy, guiding data collection toward the weakest directions, and uses in-view and cross-view checks plus the minimum detectable deviation to define the smallest fault detectable in each direction.
The framework combines a neural-network-enhanced robust wrench observer with a hybrid pose-force controller for physical human-robot interaction (pHRI), enabling end-effector pose and force tracking without force/torque sensors. The neural network compensates for system friction and unmodeled dynamics to improve observer performance, and task-null-space decomposition separates joint velocities and reduces the effect of human disturbance on the observer. Experimental results show that the framework achieves robust and accurate wrench estimation while maintaining reliable pose and force tracking, and can be applied to various robotic arm platforms without force/torque sensors.
The framework uses a meta-learning-trained neural network to recognize two human-robot interaction modes, active guidance and passive following, without manually setting switching thresholds. The passive mode recognition rate rose from 90.1% to 99.1%, and the active mode from 79.8% to 97.1%. An adaptive meta-learning controller can compensate for human disturbances and adapt online to varying interaction forces. A multi-trajectory meta-learning algorithm jointly learns an environment basis representation and force-related parameters to handle unmodeled dynamics. Experiments verified its robust control and reliable mode recognition performance.
This study proposes a coupled visual-motor remapping method that pairs panoramic azimuth warping with its inverse upper-limb yaw transformation for first-person teleoperation of humanoid robots, to extend the operator's reach to side and rear targets. The prototype implements two mapping modes: Global Zoom (GZ) and Focus+Context (F+C).