The arXiv paper "Goodhart in the Planner" measured how sampling-based control with rewards learned from human preferences performs on the true objective: on HalfCheetah, Hopper, and Meta-World's drawer-open.
LACART combines lookahead-based control barrier functions (CBFs), a relaxed control Lyapunov function (CLF), and quadratic programming (QP) to modify nominal vehicle commands in a unified optimization layer, handling both circular obstacles and track boundaries.
DKWBC proposes a dynamics-informed whole-body control framework that uses inverse-dynamics feedforward and joint PD feedback to supply tracking torques, while a residual policy learns the joint-torque corrections needed for balance. The authors state that this structure explicitly separates motion-related torque demands, simplifying the policy learning task, and that they validated in simulation and on real hardware that a single balancing policy can generalize across multiple motion datasets, maintaining tracking and balance under ground support, added payload, and physical interaction. The project homepage has been made public.
The paper proposes a lightweight path-aware frontier re-ranking method for LiDAR UAV autonomous exploration: a lightweight utility function first retains a small top-N set of frontier candidates, then time-bounded RRT* planning estimates their feasible path costs for re-ranking, and the path for the selected frontier is reused for navigation to reduce redundant global planning.
The study deploys a cucumber length-measurement pipeline based on YOLO26n detection, prompt-based segmentation, and medial arc spline estimation to NVIDIA Jetson Orin Nano and Raspberry Pi 4, and compares both against an RTX 3050 Ti workstation.
The research team proposed a two-component optoelectronic soft suction cup designed for kiwifruit that uses optical signals to estimate interaction forces with real fruit within known contact intervals. On fruit not used in training, it achieved grouped cross-validation R²=0.968±0.005 (MAE=0.138 N), with a median per-fruit held-out R² of 0.973 (MAE=0.134 N).
The research proposes having GPT-6 Astra read a robot's base policy weights and edit only 65 rows of token embeddings, raising the success rate on π0.5's 765 held-out paraphrased instructions from 86.3% to 92.0% (48 repairs, 4 regressions, paired p=1.31×10⁻¹⁰), all without gradient updates or new demonstrations.
This research proposes an inference-time obstacle-guidance framework for pretrained relative-action diffusion policies. It rolls out cumulative relative actions into a workspace trajectory, evaluates obstacle interactions along the trajectory with an SDF-based obstacle model, and during denoising propagates each obstacle response to all contributing action increments. In real-robot experiments across seven obstacle-constrained scenarios spanning four manipulation tasks, the method achieved 105/140 (75.0%) collision-free task successes, compared with 8/140 (5.7%) for absolute-action pointwise guidance. Ablations show that the temporally coupled form of relative actions yields larger gains under both obstacle models.
MDF++ is a visuo-tactile world action model that uses a diffusion backbone to learn latent representations of multisensory observations end-to-end, addressing the problem that existing world action models rely on frozen pretrained encoders whose representations are not well suited for control. The authors propose a multi-task learning strategy that leverages cross-modal correlations to learn task-relevant representations, avoiding the feature collapse caused by naive joint training. On contact-rich manipulation tasks including real-world bolt screwing and nut mating, it outperforms world action models built on frozen pretrained representations by up to 50%.
SUGARv2 proposes a unified framework for learning diverse visual humanoid loco-manipulation skills through interaction reference synthesis and a shared training pipeline. The method adapts reusable task templates to different spatial configurations to synthesize coordinated robot and object motion references at low cost, and uses reference-based imitation as a common learning objective across tasks, first training privileged teachers and then training visual policies through teacher distillation and reinforcement learning.
The authors propose TPS-A*, a lattice planner whose edges are kinematically feasible trajectories produced by parameterized trajectory generators, with certified collision checking performed in trajectory parameter space. A grid precomputed per footprint needs only a single pass over local obstacles to return a certified free distance for every candidate trajectory, making the per-node collision cost independent of footprint complexity and trajectory count.
The study proposes a planner-verified counterfactual (CF) test: holding the grasp pose fixed, a solid obstacle is inserted into the approach corridor and its points are added to the model input, after which the planner re-derives the label, flipping it from feasible to infeasible. Across eight predictors spanning point clouds, voxels, depth, and object tokens, CF-pair AUROC falls from 0.920-0.996 in natural scenes to near chance (0.470-0.504), and this flip holds across four planners, two robotic arms, and a real-object obstacle.
The authors propose A2-MEAS, an embodied multi-agent framework that uses code as a scaling interface, combining corrective language, spatial evidence, execution history, and action proposals into correction codes; during training, MetaBrain provides supervision and runtime guidance.
Why it matters: The framework uses code as a scaling interface, injects corrective language and execution history into the policy, and provides quantitative comparisons across benchmarks and real-robot tasks.
The paper proposes a safety-critical control framework for nonholonomic multi-robot systems that lets each robot circumnavigate a stationary target at a different orbital radius while maintaining spacing from neighbors and avoiding collisions. Each robot measures only its distance to the target and exchanges information only with its two adjacent robots; it drives the radius to the target value via acceleration, balances the formation via turn-rate, and uses a higher-order control barrier function quadratic program (HOCBF-QP) to simultaneously constrain target distance, geofence, and inter-robot safety.
To address the inaccuracy of conventional frequency-domain readout caused by long interconnects and environmental noise in robotic tactile sensing, this study proposes a time-domain optimization readout method that determines resistance values by fitting a time-domain observation model and using the full waveform rather than only spectral amplitudes. In simulations of a 16×16 array, the method reduced the mean readout error by 20% compared with the baseline; validation over 1,000 trials showed reductions of 24.6% in mean error and 77.6% in standard deviation, and the improvement was also confirmed in cross-point array simulations emulating tactile sensor resistance distributions.
The paper formalizes the attribution problem of inconsistencies between scans and models in LiDAR-inertial odometry and BIM fusion as fault identification over three fault classes, and derives a geometric criterion for determining whether attribution is feasible. On cases that are detectable but not attributable in a two-dimensional study, the method based on this criterion had a conditional error rate of 0.83%, compared with 31.67% for a baseline using uncertainty-based abstention; across 5,600 attributable cases, it produced 5,574 unique attributions with no erroneous attributions.
FastFunnels proposes a hierarchical multi-robot navigation method: a high-level reinforcement learning policy generates deformable elliptical envelopes in real time, while a low-level distributed model predictive control (DMPC) layer keeps robots within the envelopes and avoids collisions with each other.
The authors present RevoSim, a multimodal tactile simulator for the BrainCo Revo3 dexterous hand that jointly models 285 piezoresistive taxels and 5 visuotactile fingertips. The simulator can synchronously generate pressure, depth, RGB, and marker-displacement observations; on a single RTX 5090, its depth-conditioned marker-displacement configuration supports up to 4096 parallel environments. The authors trained in-hand manipulation and grasping policies entirely in simulation, laying a foundation for future sim-to-real transfer.
HOC2R introduces a hand-object contact trajectory extractor that combines frozen DINOv3 and SAM 2.1 features with learned heads to perform hand-object association and contact localization, extracting robot-usable motion cues from human videos.
PRISM represents reusable skills as parameter-level primitives, storing each skill family as a dedicated low-rank action-expert adapter, and combines and invokes them at inference via a block-wise VLM-conditioned router. On RoboCasa365, PRISM achieves a 56.43% atomic skill success rate and a 20.00% success rate on long-horizon composite tasks, compared with 17.86% and 17.50% for standard two-stage π0.5-LoRA fine-tuning.
The authors propose HomeLoop, a simulation framework and benchmark built on Isaac Sim that integrates robot bodies, networked devices and wearables, simulated residents, faulty sensors, and hazards that evolve over time within a single environment. The benchmark contains 20 tasks across 5 categories, where the same alert can have different causes. A runtime monitor uses temporal logic to score both the final outcome and process actions, and each action has a preset cost.
FATHOM proposes a world-model-guided diffusion action-trajectory selection framework for relative tracking of autonomous underwater vehicles (AUVs) under sparse active light markers. The method first fuses marker observations, inertial measurements, action history, and temporal dynamics using a multimodal task-supervised recurrent state-space model to form a predictive tracking belief; a diffusion action prior then generates multiple multi-thruster trajectories, and the best candidate is selected under a receding-horizon objective and continuously updated based on onboard observations.
Lodos is a simulation framework designed from scratch, using new broadphase and narrowphase collision detection and contact dynamics computation to specifically handle mesh-mesh collisions between detailed meshes. The authors state that its physical fidelity is comparable to existing mature simulators, and that in mesh-mesh collision scenarios its throughput is about an order of magnitude higher; the source code has been released. The work targets contact-rich scenarios in large-scale parallel rigid-body simulation that are limited by approximating collision geometry with primitives such as spheres, capsules, cylinders, and boxes.
ActPrune proposes a training-free visual token pruning method that moves the pruning decision ahead of the vision encoder and uses the action chunk predicted at the previous control step as a geometric prior. In the fixed third-person view, the method retains image patches near the projected end-effector trajectory along with semantic anchors; in the wrist view, it prunes patches that remain unchanged after compensating for action-induced camera motion.
The authors propose an object-centric reinforcement learning framework that treats each part as an agent and learns a shared policy for continuous assembly control. A graph neural network represents the parts, their target poses, and their spatial relationships, and predicts translational and angular velocities for each component; per-agent PPO with a centralized graph critic learns motion and temporal coordination within a single feedback policy. The study validates the approach on planar Tangram assembly, using procedurally generated target configurations and randomized initial layouts to support training and evaluation, and demonstrates parallel assembly in simulation and transfer to sequential execution on a real robot.
ForPC is a force-aware whole-body control framework that regulates a humanoid robot's cart-pushing speed through two-stage reinforcement learning. The method first infers force latent variables and state latent variables from onboard observations without external force sensors, then uses a residual policy to coordinate arm, waist, and leg movements; in simulation, cart-pushing speed RMSE is reduced by up to 68.2% compared with a locomotion-only control baseline, and lateral error on multi-curve routes is reduced by up to 74.2%.
CS-Pi proposes a "sample-project-renoise" flow-matching policy sampler that satisfies constraints absent from demonstrations at inference time while keeping actions plausible under the policy. The method projects each sample onto the nearest safe action, then adds noise back according to the size of the projection displacement and reruns the original policy until the sample is safe. On a 7-DoF shelf grasping task it achieves an 81% safe success rate, and 55% in zero-shot deployment on Franka, versus 55% and 40% for projection alone. The paper has been accepted to CoRL 2026.
The Meituan Robotics team has released Meituan-Robotics-0 (MR0), a vision-language-action foundation model and embodied-agent system for fixed-base desktop manipulation. Its training data includes 147 million vision-language samples and 15,060 hours of robot trajectories covering 12 embodiments.
Apple's research team published the paper "Perceptive Humanoid Motion from Simulation" on arXiv, unveiling HALO, an autonomous humanoid robot control system, and validating it on the Unitree G1.
Why it matters: Apple's HALO splits task orchestration, motion generation, and tracking into three layers, and reports comparative results of 91% in unfamiliar environments and 70% on composite tasks, offering a look at its layered design and data synthesis pipeline.
SuperMimic proposes a humanoid robot motion-tracking framework that uses a differentiable world model with 32-step backpropagation rollouts to reduce training compute demands when the motion library is large. The framework models physical laws with a deterministic incremental model and uses real states as the starting point for imagined rollouts. The original is an in-depth analysis; the body provides only a summary.
Long-WAM learns visual dynamics through autoregressive video pretraining and uses asymmetric causal attention to connect video and action experts, enabling long-horizon decision-making based on historical information. Its approach predicts future frames in latent space, combines historical information to generate action chunks, and supports streaming encoding and asynchronous execution.
Tsinghua University and other institutions have proposed the ART (Agentic Robot with Tool-use) framework, which can fine-tune any end-to-end VLA model into an agent capable of plug-and-play use of external tools. The method uses 30,000 tool-call trajectories and two-stage LoRA fine-tuning, achieving average success rates of about 75% in LIBERO simulation and about 62% on a real Astribot S1 robot, an improvement of about 20% over baselines such as OpenVLA and π0.
Amodal3R at ICCV 2025 performs occlusion reasoning and geometry generation directly in 3D latent space, using visible-region masks and occluded-region masks to preserve real observations and infer occluded geometry and appearance.
WorldFoundry proposes open-source infrastructure for world intelligence, intended for reproducible inference and evaluation across video, 3D and 4D, embodied action, and interactive world systems. The work defines a generative world model as a predictive dynamical system that maintains a persistent, history-conditioned state representation, responds to actions or interventions, and generates temporally coherent trajectories. Its catalog includes 270 model entries, 74 benchmarks, and 62 metric entries, and it records execution readiness and evaluation conclusions separately through manifests, artifact contracts, and evidence-aware scorecards.
Apple released the HALO humanoid whole-body motion control system, which procedurally generates 1,000 virtual residential scenes and synthesizes nearly 300 hours of paired data, enabling the Unitree G1 to master skills such as sitting on chairs, turning sideways to pass through a half-open door, crouching under a table, and pushing boxes.
Why it matters: Apple's HALO uses simulation-synthesized data to train humanoid whole-body skills, transfers them zero-shot to real scenes, and ablation experiments show the contribution magnitude of each module.
Yuance Future has released Anchor 1, a whole-body control model positioned as the System 0 cerebellum layer within a whole-body intelligence architecture. It uses five anchor points—the pelvis, left and right wrists, and left and right ankles—as commands and does not rely on online retargeting.
Why it matters: The original article provides quantitative comparisons of the whole-body control model's real-machine error and teleoperation collection efficiency, which can be used to assess how the cerebellum layer's precision affects the success rate of upper-layer VLA tasks.
A team from Northwestern, CMU, UNC Chapel Hill and Amazon, in the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," defines world modeling as estimating the state transition p(s′|s,a) and builds WOVEN, a training and evaluation base of 36K questions.
Why it matters: The paper breaks world modeling into independently intervenable state-transition tasks and provides reusable rules for selecting training data, offering reference value for multimodal model training.
The Hong Kong University of Science and Technology, Beijing Academy of Artificial Intelligence, Peking University and other institutions jointly proposed DexRoam, which uses first-person human whole-body demonstrations to train mobile dual-arm dexterous manipulation policies. The work has been accepted to CoRL 2026. The system narrows the human-robot motion distribution gap through three stages: embodiment alignment, action semantic alignment, and temporal alignment, reducing MMD and SWD metrics by an average of 41.7% and 31.1%, respectively.
miHoYo, together with Tsinghua University, the University of Hong Kong, the National University of Singapore, and other institutions, has open-sourced the GameScaling series of general game models, covering four size tiers—0.8B, 2B, 9B, and 27B—for a total of eight models. The weights and inference code are released under the Apache 2.0 license.
Researchers at ETH Zurich changed how muscle connects to the frame, enabling a 1.5-centimeter robot to jump 5 millimeters with a single muscle contraction. They designed the muscle together with a flexible 3D-printed scaffold; tiny grooves on the scaffold help muscle fibers align neatly and provide firm anchor points for force.