LANav targets the open-vocabulary object goal navigation (OVON) task, in which a robot must find a target in an unfamiliar indoor environment using only its current first-person image. Its approach is to have the robot stop repeatedly reviewing past frames and instead organize its memory to support navigation decisions.
RLT proposes freezing the large VLA model and training only a lightweight actor-critic control head: the VLA handles visual and language understanding and coarse actions, while the small network corrects actions online during critical stages to make up the last few millimeters of precision. The two are connected through a bridging structure.
The study proposes 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 designed 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.
A digital fruit fly reconstructed from micro-CT scans achieves hexapod walking in MuJoCo; the tripod gait self-organizes from phase-oscillator phase-locking, with reflex correction rules and adhesion measurements added. Its compound-eye vision outputs 721 ommatidia per eye, and all conclusions come from complete simulations with fixed seeds.
Cortical Labs' biological computer CL1, built from living cells, was trained in a simulated 2D environment and then used to control the movement of a real robot dog. A paper from the company that has not yet been peer-reviewed shows that the neural network improved over time when continuously fed the same data, and Ferrante Neri of the Education University of Hong Kong considers this preliminary evidence of learning from limited experience. Cortical Labs says the biological computer may consume less energy than digital approaches, and its cells come from blood stem cells donated with donor consent.
Researchers developed a single control policy that lets Unitree's G1 complete ladder climbing, crawling, obstacle traversal, stepping stones, and recovery after falls by deciding the placement of nine different body parts. The robot can use its hands, elbows, knees, feet, and even pelvis for support rather than relying only on walking, achieving a 92.5% success rate on a simulated obstacle course, but how this flexibility performs in unpredictable real-world environments remains a bigger test.
Long-WAM from teams at NVIDIA, MIT, the University of Hong Kong, and UCSD, and RoboJEPA from Meta FAIR, Mila, and other institutions were submitted to arXiv on the same day, October 7, measuring the returns from adding visual history and from expanding training compute, respectively.
Zhang Changshui and Cui Sen's team at Tsinghua University propose RoboCoach, an embodied intelligence system that uses the world model CoachWorld to run closed-loop trials of long-horizon tasks in imagination, identify the first weak subtask that times out, and request targeted human teaching for it, updating only the corresponding skill expert. Under a fixed demonstration budget, the success rate rises from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX, while the uniform collection baseline reaches only 30.0% and 47.5%, respectively. On four unseen new task routes, the average success rate is 35.0%, compared with 0% for the shared-policy baseline.
Why it matters: The paper presents a complete experiment on using a world model to locate weak subtasks and improve success rates through targeted teaching; its approach to data budget allocation may be a useful reference.
Researchers from UIUC, Cold Spring Harbor Laboratory, Brown University, and the Max Planck Institute propose DSReg (Dependency-Sparsity Regularization), which can be added directly to an already-trained LeJEPA so that each learned dimension corresponds to a real-world variable, without reconstruction of observations, a decoder, or labels.
The Samsung RX Lab-China team proposes ε4P (Imperfection for Precision), which admits imperfect data into different noise intervals of flow matching according to the type of data defect, thereby turning data that is harmful when mixed directly into training into a performance gain.
Researchers from Tsinghua University and the National University of Singapore propose ParallelWorld, which adds multi-step candidate future rehearsal to high-level planning for embodied agents, raising task success rate from 56% to 83%. The method generates and verifies multiple candidate action branches in a rehearsal loop; during execution it takes only the first action of the selected trajectory, then replans based on new observations. Success rates were higher than the direct baseline across eight LIBERO-Pro settings, but adaptive search took on average about 5.8 times as long as direct execution.
DUET proposes separating a humanoid robot's manipulation policy from its balance controller and combining the two commands only at the robot interface, so learning new manipulation skills no longer requires retraining the gait controller. The authors say this removes the hidden cost of retraining the balance controller every time the manipulation policy changes; the paper, code, and a full summary are in the first comment.
A NVIDIA team, together with Simon Fraser University, proposes GATOR, which uses generative reconstruction as a base and multimodal agents for local refinement to reconstruct complete 3D objects with scene poses from casually taken photos. The method is trained only on synthetic scenes and, without any real-world fine-tuning, leads baselines on geometry, appearance, and pose metrics in real tabletop and indoor scans; without refinement, it takes about 0.5 minutes to outperform GPT-6-Astra running for 20 minutes. The authors also note that reliance on synthetic data, binding to commercial APIs, and stability under extreme poses remain open issues for deployment.
Researchers have developed a bionic robot called BeetleBot that mimics a beetle's body and limb structure and gait control, allowing it to move stably across complex terrain such as rubble, sand, and narrow crevices. Experiments show it can climb over complex obstacles and quickly recover its movement state after external impacts or being flipped over. The robot is expected to be used in post-disaster search and rescue, exploration of hazardous areas, inspection of industrial facilities, and environmental monitoring.
The authors propose the CATS method, which combines a temporal penalty with linear ramp-up to suppress action oscillations in deep reinforcement learning policies. The paper proves that the temporal penalty can bound the expected action differences between current states that share the same successor state, thereby producing a spatial smoothing effect. In simulations and real-world environments, the method is verified to significantly reduce action oscillations without degrading task performance, with very low computational overhead. The work was submitted to ICRA 2027.
GenTrack3 proposes an online multi-object tracking (MOT) framework that integrates deterministic and stochastic principles, using stochastic particle sets to compute target-to-detection matching costs in order to maintain robust tracking under uncertainty. The paper also introduces a new track-to-detection matching method that improves scalability as the number of targets grows and supports group tracking; tracking inference uses tracklets containing identifiers, states, velocities, track penalties, and track ages. The authors state that a reference implementation and baseline trackers have been released on GitHub.
The authors propose TouchSafeBench, a physically grounded benchmark built in Habitat 3.0 for evaluating the collision grounding capability of vision-language models (VLMs). It contains 2,940 simulated indoor co-presence episodes covering two scenario types, social navigation and social rearrangement, and provides synchronized multi-view RGB-D observations, top-down trajectory maps, calibrated camera metadata, and simulator contact labels.
The authors propose TopoGS-SLAM, which decouples refinement from capacity using coverage-triggered Gaussian initialization and bounded-topology optimization, turning mapping into a bounded-dimensional problem. On standard indoor benchmarks, its localization and reconstruction accuracy is not inferior to densification-based pipelines, while memory scalability and throughput improve significantly. The authors therefore conclude that in bounded-area static indoor scenes with sufficient multi-view coverage, Gaussian SLAM does not require iterative topology mutation.
GenTrack2 proposes a visual multi-object tracking method that jointly uses stochastic particle filtering and deterministic association to maintain identifier consistency under conditions where the number of targets is unknown and changes over time and the dynamics are nonlinear. The particle filter uses particle swarm optimization (PSO) to guide particles toward the modes of the state distribution, with a fitness measure that combines motion consistency, appearance similarity, and social-interaction cues with neighboring targets; deterministic association strengthens identifier consistency through a cost matrix that incorporates spatial consistency between particles and current detections, detection confidence, and track penalties.
DepTR-MOT proposes a category-agnostic DETR-based detector that embeds instance-level depth awareness into the detector through instance-level soft depth label supervision and dense depth map distillation, and outputs instance depth at inference to improve multi-object tracking under occlusion and close-proximity interactions.
The authors propose LENS, which accelerates ACEsplat-style scene memory workloads on an AMD KR260 FPGA/ARM platform, comprising an INT16 scene coordinate regression accelerator and a fixed-point ordered 3DGS renderer.
The authors propose ContactBench 1.0, which uses 36 reference tasks and more than 40 object assets to break contact-rich manipulation down into eight essential skills; the assets can be reproduced on consumer-grade 3D printers and each comes with a matching simulation model. They evaluate recent vision-only and visual-tactile policy learning methods on six benchmark tasks, and the results show that vision-only policies struggle to perceive or respond to contact events, while visual-tactile policies do not bring noticeable improvement. The benchmark will be open-sourced, providing physical and simulated assets and controlled task variations for evaluating the generalization ability of robotic systems.
The paper proposes a series-parallel humanoid leg design that combines a knee parallel mechanism, an ankle parallel mechanism, and a nonlinear knee-ankle coupling, with all actuators placed near the hip. The authors derive coupled nonlinear transmission models and validate them in simulation and on hardware. The 6 kg leg achieves an effective reflected mass of 0.37 kg at the foot end; the coupling passively assists ankle dorsiflexion to lift the foot when the knee flexes, and assists plantarflexion for push-off when the knee extends.
ArtifactArena is an open-ended evaluation platform that challenges models with hardware-software co-design: building working robots in a simulated arena and pitting them against each other. The platform guides models to improve their robot artifacts through three channels: text descriptions, physics simulator feedback, and gameplay data, and compares frontier models using Elo rankings from head-to-head tournaments. The authors are releasing the framework and tournament infrastructure for ongoing community submissions, creating a non-saturating long-term testbed.
The research team proposes DASH, a da Vinci adapter for serial-link and humanoid robots that mounts da Vinci Classic and Xi instruments onto general-purpose robots without modifying the instruments themselves. The two adapter types are compatible with robotic platforms of various load capacities, engage the instruments' native latches, and wirelessly identify inserted instruments to load the corresponding kinematics and coupling matrices. Teleoperation experiments validated the approach and its role in lowering the barrier to surgical robotics research.
The authors propose RMMBench, a mobile manipulation benchmark that requires robots to understand language instructions and execute long-horizon tasks in continuous space, with a "navigation-manipulation" task suite covering 70 typical task scenarios. Results show that leading vision-language models (VLMs) still face notable challenges in spatial localization for mobile manipulation tasks, and their spatial perception in long-horizon interactions needs strengthening. The work has been accepted to CoRL 2026.
The study proposes Embodied Agent Arena, evaluating ten VLM agents across five dimensions—geometry, spatial reasoning, affordance, task planning, and manipulation—with 1,000 cases, including the 168-case GeoProbe benchmark for camera, motion, depth, and scale estimation.
The authors propose WorldEcho, which evaluates whether action-conditioned world models truly follow commanded actions over a broader action distribution through visual integrity and SE(3) end-effector trajectory alignment. The diagnosis shows that models trained on expert data perform well on demonstrated actions but produce action-inconsistent or visually invalid generations on non-expert trajectories.
PartialBiGrasp is a dual-arm grasp generation framework that operates directly on partial point cloud observations. It implicitly learns geometric features through a convolutional occupancy network to reason about graspability, collision-free contact regions, and object thickness. Based on this, the method generates force-closure-compliant grasp pairs and uses sampling-based optimization to correct for ambiguity caused by incomplete geometry. The authors evaluated the method using analytical force-closure metrics, large-scale simulation experiments, and noisy partial point clouds on a real robot.
This paper has been withdrawn by its author, Zeyu Ni. OccamView is an object-conditioned view-selection framework for frame-budgeted active 3D Gaussian reconstruction. It builds an online object memory from open-vocabulary detections, represents unresolved local occupancy around detected objects as conservative hidden-region proxies, and evaluates candidate viewpoints with an occlusion-aware proxy-coverage score.
UCA-Flow proposes a unified condition-action modeling framework that encodes observation conditions, timestep conditions, interval conditions, and action tokens into a single sequence, which a unified Transformer learns jointly to enable accurate one-step action generation.
RAPAC-DP proposes a response-aligned pending-action compensation framework that enables diffusion and flow-based action generators to still output full action chunks aligned with the handover moment under cloud inference latency. The method uses a lightweight pending-action encoder and LoRA adapters to learn the compensation while keeping the base policy frozen, and constructs delay-conditioned training samples from delay-free demonstrations, requiring neither an explicit dynamics model nor additional delay data.
The paper proposes Environmental Constraint Exploitation (ECE), a method that deliberately makes contact with the environment to reduce the dimensionality and computational complexity of motion planning. The authors integrate ECE into an RRT-based planner, biasing exploration toward task-relevant regions and using contact to reduce uncertainty and improve execution robustness, and demonstrate practical results in real-world scenarios. The work consolidates and extends prior research, showing that ECE can simplify motion planning and enhance adaptability in complex environments.
The paper proposes Semantic Anchoring, a plug-and-play method that anchors VLA action representations to a semantic manifold and splits them into a shared semantic channel and a private channel; both are discarded at inference, leaving the deployed model unchanged.
The authors propose the Language Sensor Model (LSM), which maps each natural language description and its scene-graph context to a multimodal distribution, using mixture weights to represent referential ambiguity and component covariances to represent spatial uncertainty.
This survey examines three roles of 3D generation in embodied AI and robotic simulation: as a data generator producing simulation-ready assets with articulated structure, physical properties, and deformable content; as a builder of interactive, task-oriented scenes for simulation environments; and as a sim-to-real bridge supporting digital twin reconstruction, data augmentation, and synthetic demonstrations. The authors note that the field is shifting from visual realism toward interactive usability, and they list key bottlenecks including limited physical annotation, the disconnect between geometric quality and physical validity, fragmented evaluation, and the sim-to-real gap.
This paper proposes estimating the coefficients of a compact Geometric Variable Strain (GVS) parameterization within a factor graph framework, and uses a kinematic factor derived from the Magnus expansion of the strain field to link the GVS strain coefficients with backbone pose variables, thereby reconstructing the shape of a continuum manipulator. The method is compared against a Gaussian process regression baseline on simulated and experimental data, achieving higher accuracy in partial and sparse measurement scenarios while maintaining computational efficiency. The state vector is compact and can be used directly for model-based control.
The authors propose ENAP (Emergent Neural Automaton Policy), a neuro-symbolic imitation learning framework that uses adaptive clustering and an extension of the L* algorithm to learn a Mealy state machine from demonstrations as an interpretable high-level planner, which then guides a low-level reactive residual network to learn continuous control via behavior cloning.
The authors propose VORL-EXPLORE, a hybrid learning and planning framework that uses execution fidelity to estimate local navigability and execution reliability, coupling it into a Voronoi allocation objective with inter-robot repulsion and arbitrating during navigation between global A* guidance and a reactive reinforcement learning policy. The fidelity model is recalibrated using self-supervised pseudo-labels generated from recent progress and safety outcomes. Experiments on randomized dynamic grids and an ablation under severe congestion show higher success rates, shorter exploration paths, and lower inter-robot coverage overlap, along with a deployment example in a Gazebo factory scenario; the code is open-sourced.
The authors propose Consistent Action-Response Execution (CARE), which grounds local control in executable-action-indexed joint human-robot transitions, so that the same executable action simultaneously determines operator response prediction, joint human-robot rollout, and candidate selection. On training and held-out maps, CARE maintains near-saturated performance when the indicated route remains locally passable, and achieves the highest success rate and service availability when the operator can continue but the robot must detour.