A Tsinghua University research team proposes SimpleTouch, which adds a tactile expert to the pretrained VLA π0.5 and uses single-stage training with only 50 demonstrations per task, without additional tactile policy pretraining or a separate alignment stage.
An ECCV'26 paper proposes MV2GF, a multi-view pedestrian detection method that uses the multi-view features and 3D point maps of the pretrained visual geometry foundation model DA3 to aggregate BEV features instead of perspective projection. Under the setting where training and test camera configurations differ, MV2GF achieves MODA on GMVD-D, MVPerception, and Wildtrack that is 4.6, 4.7, and 2.2 points higher than the previous best, respectively.
EmbodiedLab is a benchmark that requires robots to "acquire knowledge through action," covering 30 quantitative science experiments across 5 disciplines on 14 apparatus platforms. The parameters of each instance are resampled, so results cannot be known in advance but can be scored precisely.
The authors propose the world action model Video Prediction Policy 2 (VPP2), which uses event-level pretraining on large-scale manipulation video, distills the video model into a single-step visual planner, and connects an action module via a MoT architecture to learn implicit inverse dynamics.
The authors propose the IG-VLA framework, which lets a vision-language-action model imagine task-relevant future scene evolution in visual representation space and internalize the reasoning results into a compact Scene Gist Token via Scene Gist Memory, thereby skipping explicit future imagination at inference time.
The authors propose JRDB-AVR, an active visual reasoning benchmark built on real JRDB robotics data: after receiving a visual reasoning question, an embodied agent must request limited observations by timestamp and viewing angle, and evaluation examines both the final answer and the visual evidence supporting it.
FlatVPR proposes a plug-and-play geometric linear residual adapter that applies a residual transformation to features from foundation models such as DINOv2-ViT-S/14 and uses a Pullback Flatness Loss to reduce manifold curvature, so that descriptors between adjacent anchors can be reconstructed by linear interpolation. On the NCLT dataset, the adapter still delivers clear performance gains even when anchor spacing reaches 100 m and severe seasonal changes are present. Map construction is incorporated into an EM framework, split into a continuous M-step for manifold adaptation and an E-step for the anchor selection criterion.
An ICLR 2026 paper re-examines the question of whether differentiable simulators can give better policy gradients, arguing that bias from discontinuous dynamics is not the main obstacle. The authors propose DDCG, a lightweight test that switches estimators in nonsmooth regions; it requires only one hyperparameter and remains reliable with small samples. They also propose IVW-H, a per-step inverse-variance implementation that stabilizes variance in differentiable robot control tasks without explicit discontinuity detection. The paper concludes that estimator switching can improve robustness in controlled experiments, while in real deployment variance control is often more critical.
Prateeth Rao and Sachit Rao propose a geometry-aware epipolar correspondence graph method that uses epipolar constraints to guide the construction and pruning of correspondence graphs, then a graph learning module generates a geometric context vector, which is passed through an MLP to output rotation and translation for camera pose estimation.
The authors propose ActECC, which addresses the difficulty of determining the valid range of action chunks and scheduling replanning requests in edge-cloud collaborative VLA inference through kinematics-guided usable range assessment and delay-aware replanning scheduling. Experiments show it reduces per-action latency by up to 27.4%, cuts cloud calls by up to 45.1%, and lowers average execution error.
A research team proposed a simulation-to-real robot learning framework for agricultural reaching and grasping, building a MuJoCo-based agricultural environment with diverse tomato plant structures and collecting expert demonstrations via joystick teleoperation. Two pretrained generalist policies, Octo and π0.5, were first adapted with simulated demonstrations, then fine-tuned with real demonstrations on a Franka Research 3 robotic arm. Results showed that π0.5 scored higher than Octo on task performance and degraded less under distribution shifts such as viewpoint and spatial offsets; the study also examined the effect of proprioceptive information on policy performance.
CLARO is a VLM agent that orchestrates skills such as open-vocabulary object detection, segmentation, 6-DoF grasping, and non-prehensile pushing to locate and retrieve a language-specified target in cluttered environments. It records past actions to support long-horizon execution and outputs grasp poses or 3D waypoints, enabling transfer to different robot arms without task-specific training. In a dynamic simulation environment, CLARO achieved an average success rate of 89.1%, a relative improvement of over 16% on the strongest competing method in highly cluttered scenes; on a real robot arm, its success rate was 88.8%.
The authors propose the PlayPolicy framework, in which a robot first learns a play policy from minutes of human play data through behavior cloning, then deploys it autonomously to collect hours of interaction data to train a world model; at inference time, this policy proposes candidate action sequences for model-predictive control (MPC).
The authors propose an inspection planning framework that replaces geometric visibility checks with a learned camera information metric based on a classifier variational autoencoder (VAE), to address inspection failures caused by drastic orbital illumination changes. The method implements a guidance, navigation, and control architecture in a photorealistic physics simulator and tests multiple orbital illumination conditions and viewpoints: average brightness reaches 0.57, with 81.6% of observations above 0.25, compared with 0.02 (3.0%) for geometric planning with illumination bands and 0.04 (6.0%) for pure geometric planning.
This study formulates upper-limb functional mobility estimation as an active perception problem, parameterizing reductions in healthy joint limits with impairment coefficients and integrating this into a human model used during robot interaction. The authors propose an online active perception pipeline in which the robot actively probes the user through repeated handover requests while progressively updating a Bayesian belief over the impairment coefficients. Simulation studies and a proof of concept on a robotic platform show that the pipeline can recover the impairment coefficients with relatively high accuracy in real human-robot handovers.
Researchers propose the CapPro framework, which lets a vision-language-action (VLA) policy adapt to previously unseen manipulation tasks using only a single human demonstration video, without robot demonstrations or fine-tuning for the target task. The framework has two components: a phase-aware human video representation module, which learns task-relevant query representations through visual token selection and phase-caption supervision, and a video-conditioned adaptation module, which injects the compact video and query representations into the action-conditioning context of a pretrained VLA policy.
OneDemo proposes a reinforcement learning framework that edits a single successful demonstration into trajectories suited to new task configurations; the learned editor serves both as a task policy and as a reusable demonstration generator. On 15 Meta-World tasks, it reaches a 92.7% success rate using an average of 53.7k environment interactions; on 9 DexMimicGen tasks, it generates demonstrations with an average success rate of 98.4%, and BC-RNN policies trained on the generated data reach an average success rate of 99.6% at the same data scale.
DiVeR estimates decision criticality from the dispersion of sampled action representations and uses it to reweight verifier learning, focusing learning on states where action selection matters most, without requiring step-level annotations or additional environment interaction. In experiments on LIBERO, RoboCasa, and a Franka Research 3 real robot, DiVeR consistently improves task success rates by guiding action selection with a more effective verifier, and verifier inference overhead is negligible.
CognitiveReality builds a semantically indexed Gaussian-TSDF map from a robot's RGB-D stream, shared by a VR operator and a tool-calling language agent. A single mapping program adapts to different platforms through configuration alone, accepting poses from robot SLAM, joint kinematics, motion capture, or visual tracking, and handles localization outages with shadow tracking and keyframe-anchored PnP, maintaining open-vocabulary instance identities and per-object quality at 2 Hz.
DGT-Map proposes a self-supervised framework that learns global, directional, and vehicle-conditioned traversability costmaps from RGB-D observations and locomotion signals. It uses a shared multi-task backbone to learn common terrain features across vehicles, while vehicle-specific prediction heads preserve platform-dependent responses. At inference, it outputs a heading-indexed costmap for use by direction-aware planners.
StepTrigger is a contact-state-triggered backdoor attack on VLM-powered legged robots. The trigger signal comes from the foot-ground contact pressure patterns produced when a Unitree Go1 quadruped walks across dense terrain, rather than from prompts or visible markers.
H2RBench is a Real2Sim benchmark for evaluating human-to-robot (H2R) transfer methods. It establishes a standardized protocol based on real human video demonstrations and simulated robot demonstrations, covering four categories of manipulation tasks.
The authors propose Expert-Play Contouring Control (EPCC), which combines slow expert demonstrations with fast, non-expert play to achieve execution speeds faster than the demonstrations. Expert demonstrations train a latent trajectory generator whose predictions are reparameterized into a time-independent contour of task progress, while play data trains a world model of fast-action outcomes. At deployment, the planner optimizes actions along this contour and penalizes deviation from it.
The paper proposes a path-tracking algorithm for articulated vehicles with a single trailer. It applies a lateral control law at the trailer level and makes short-horizon predictive adjustments to the tractor steering angle, so that both forward and reverse motion converge stably to the target path. The method is geometry-based and requires no per-vehicle calibration or training. In a high-fidelity physics simulation, the steady-state or final lateral error on straight-line, circular-arc, and reverse docking trajectories was 0.02–0.05 m. The authors tested it on a small-scale tractor-trailer robot, with the controller running on a low-cost microcontroller; under about 200 ms of pose-detection latency, the lateral error in repeated reverse docking was below 13 mm.
The authors propose VLN on the Fly, an onboard vision-language navigation stack that keeps instruction grounding, planning, and control as separately inspectable stages: a quantized VLM grounds the instruction to a coarse image cell, depth lifts it to a 3D goal, a fast B-spline planner generates a feasible trajectory, and a pretrained reinforcement learning policy tracks the trajectory and outputs motor commands across multiple quadrotors.
ForceDelta-VLA proposes a correction-distillation framework that constructs an explicit force-correction target from the force-conditioned and force-agnostic predictions of a frozen teacher model, and adds a separate delay-correction target to handle reference-action mismatch. Across nine single-arm and bimanual contact-rich tasks, the system achieves a mean success rate of 82.2%, compared with 54.4% for the ForceVLA baseline and 70.6% for direct execution of the Stage-1 temporal teacher. Compared with ForceVLA, the full system reduces the mean peak contact force of successful trials by about 26% on both platforms.
X-WBC is a cross-embodiment humanoid whole-body control foundation framework that separates relatively shared human motion semantics from embodiment-specific physical execution. It uses human-centered command tokens to align full human motion, robot reference motion, and sparse VR observations; a causal Transformer learns reusable temporal structure from mixed multi-robot rollouts, and lightweight robot-specific modules map the shared representation to each robot's proprioception and action space.
Why it matters: The paper proposes a cross-embodiment humanoid whole-body control framework that decouples shared human motion semantics from each embodiment's execution, and validates joint training on nine simulated embodiments and four real robots.
Researchers propose a "motion communication" method that encodes arbitrary short messages (such as an agent's current intent) as noise into the actions of any pre-trained policy, making robot motion detectable via remote vision or motion capture without degrading policy performance. This physical channel does not rely on wireless communication and requires no additional hardware or direct link to the robot. On real robots running at 50 Hz, four robots jointly recovered an 8-bit message at a total rate of 0.8 bits/s, with validation completed in both simulated environments and real-robot deployment.
The authors annotated each failure of OpenVLA-OFT and pi0.5 on out-of-distribution LIBERO-90 tasks, identifying four failure modes, of which only execution misalignment can be fixed by a single retrospective language correction.
This work proposes using sample-based model predictive control (SMPC) in simulation as an automated expert to generate large-scale offline datasets, then training an off-policy reinforcement learning agent with purely sparse task rewards, eliminating the need for manual dense reward design. Combined with a low-level dynamic stability controller, the high-level agent's learned policies can exceed the original optimal control teacher on task objectives. The authors completed sim-to-real deployment on an arm-equipped Spot quadruped and a G1 humanoid; the paper is the accepted version for CoRL 2026.
The SHRIMP system lets users automatically generate hierarchical robot primitive plans using natural language, and iteratively revise those plans through re-prompting and explicit correction. After each revision, users can validate the plan in simulation and, once satisfied, execute it on the physical robot. A user study with 35 participants on tabletop kitchen task planning showed that SHRIMP improved users' perceived control and robot transparency.
The study proposes DSMS, a contact-implicit, direct-simulation-based multiple shooting framework that converts kinematically feasible reference motions into dynamically feasible whole-body trajectories. It handles contact, friction, collision, and joint limits within a nonlinear program without presetting a contact schedule. Compared with existing retargeting methods, DSMS accelerates motion-imitation reinforcement learning training and yields policies with high success rates and low tracking error. The authors achieved zero-shot sim-to-real transfer on the Unitree G1, completing walking, jogging, contact-rich crawling, and a 180-degree jump-turn.
The study proposes Staged Multi-Agent Training (SMAT), a method that jointly trains a musculoskeletal human agent and a bilateral hip exoskeleton control policy in physics simulation to learn an assistance policy that accounts for human-device co-adaptation. The policy was transferred to a physical hip exoskeleton without subject-specific retraining and tested in eight healthy adults; indirect calorimetry showed that active assistance reduced net metabolic rate by 19.7% relative to the passive device (p < 0.001), the hip mechanical positive-power ratio was 0.984, and effective assistance was maintained at walking speeds and terrains beyond the training conditions.
The paper proposes a robot-agnostic compliant control framework that separates ROS controller boilerplate from control-law implementation through a plugin-based execution architecture, extending the ROS control ecosystem. The framework uses Pinocchio to compute robot kinematics and dynamics from URDF models, and its standardized joint and Cartesian command interfaces support motion references, stiffness and damping gains, nullspace targets, desired interaction wrenches, and independent feedforward inputs.
This paper proposes a ZMP modeling method for quadruped robots carrying payloads via a passive arm, relating the stiffness, damping, and payload mass of the passive interface to the ZMP response, and notes that lightly damped passive modes may overlap with gait harmonics.
LaWAM proposes a Latent World Action Model that exposes predictive dynamics to robot policies through compact latent visual subgoals rather than reconstructed future video. Its core LaWM trains a latent action model in the latent space of a pretrained vision foundation model and reuses its forward decoder to predict future observation features.
GeoHAT is an end-to-end diffusion policy framework that uses a lightweight Fourier spatial encoder to map per-pixel 3D coordinates into geometric tokens, then injects them into vision foundation model features via depth-validity-gated fusion, without requiring an additional 3D vision backbone.
This study proposes a spatiotemporal informative planning method based on model predictive path integral (MPPI) control for unmanned surface vehicles (ASVs) to search for and capture multiple drifting targets. The planner couples target drift prediction with exploration objectives through a cost design, optimizes continuous trajectories over long horizons, and directly generates kinematic commands; in the interception stage, pure pursuit guidance is used to achieve physical capture. In simulation, the method captured 18% more targets than the strongest baseline, and the approach was validated in real-boat trials.
The authors propose PRISM, which trains a lightweight action head on features from a frozen latent world model encoder to obtain a goal-conditioned action prior, then uses its mean and predicted uncertainty to guide action search, with the world model evaluating the consequences of actions. Experiments on four simulated control tasks and two real-robot platforms show that the lightweight action head improves planning performance, and that the same fusion interface can accept a diffusion prior over the same frozen features without modification.
The authors propose MRO-GWM, an object-centric Gaussian world model for learning action-conditional dynamics of 3D rigid objects. The model represents arbitrary shapes and multi-object scenes with Gaussians in an object-centric frame, and uses a temporal-spatial Transformer to predict rigid body motion from past Gaussians and future actions; training is based on multi-view reconstruction, which requires handling partial observations caused by occlusion.