Professor Lan Xuguang's team at Xi'an Jiaotong University has proposed AffordSim, which connects open-vocabulary 3D affordance prediction to simulation trajectory generation. Starting from a natural language task description, it locates an object's functional regions, then combines cuRobo motion planning to generate executable manipulation demonstrations.
Gamma Intelligence, TeleAI and Unitree have jointly launched LocoManiLab, a unified evaluation framework for humanoid robot mobile manipulation. It covers five types of home spaces, 13 scenario instances and 234 tasks, with the Unitree G1 and Dex3 dexterous hands as the unified test platform.
A team at the University of California San Diego proposed EgoThumb in an arXiv preprint, a wearable robotic thumb driven by a palm-side wide-angle camera and an imitation learning model that can autonomously grasp and stabilize objects, allowing users to complete tasks such as unscrewing bottle caps and stirring tea with one hand.
A paper proposes that how tactile data is represented may matter more than resolution. The authors pretrained a tactile encoder, LingBot-Touch, on various sensors ranging from 3x3 glove arrays to 240x240 optical fingertips, and tested it with 4 strategies on 26 real tasks.
VOMMI is a portable demonstration collection and learning framework that connects portable RGB demonstrations to VLA post-training through offline trajectory reconstruction and online visual-motion conditioning. R2-VO refines trajectories using sparse geometric anchors, reducing the average absolute trajectory error of body and hand streams by 24.6%; a policy post-trained only on portable demonstrations has an 18.2% lower base velocity error than a policy trained on robot-collected demonstrations, with comparable end-effector translation accuracy.
The study constructs a geometric visual pretraining dataset containing only a plane, an object, and a cube hand, with trajectories generated automatically and scenes without textures or backgrounds, so that the policy is first exposed to the geometric relationship between the hand and the object. The authors evaluate ACT on three simulated robots with five manipulation tasks each and on three real-world robot tasks; in most combinations, fine-tuning from the geometric prior yields higher success rates in early training than training from scratch, and requires only a small number of task demonstrations. The results indicate that highly simplified geometric scenes can also provide an initialization that transfers across robots and to real-world tasks.
The authors propose LAG-Fusion, a latency-aware guidance fusion framework for asynchronous multimodal diffusion policy composition that allows modality-specific policies to run at their own inference rates and provide denoising guidance whenever output is available. To keep asynchronous composition consistent, the authors derive a reference-frame rebasing rule for diffusion variables under relative action representations, so that delayed guidance is aligned before fusion. In contact-rich manipulation, the authors compose a low-frequency vision policy with a high-frequency force policy; experiments with heterogeneous modality latency show better responsiveness and task performance than synchronous fusion and a purpose-designed force-aware baseline.
NoContactNoWorries proposes a Transformer-based multimodal framework that fuses RGB-D vision with robot proprioception to infer binary contact states for hand-object interaction as a pseudo-tactile signal. The authors train a single contact prediction model on multiple objects and verify that this inferred signal can support downstream reinforcement learning agents in completing in-hand object reorientation, and that it generalizes to new objects. Both simulation and real-robot experiments validate the feasibility of inferring contact from vision and proprioception. The paper has been accepted to IROS 2026.
The authors propose DexTeleop-0, a bimanual dexterous teleoperation framework that uses fingertip tactile feedback to overlay robot-side contact corrections on human tracking commands, mitigating kinematic mapping errors and the inability of vision alone to regulate contact force. The method computes localized corrections in real time via contact-stiffness approximation and the kinematic Jacobian, and adds per-contact force tracking, soft regularization of predicted contact moments, and bounded hand-joint residuals.
The study proposes Learning-Augmented Robotic Automation, a hybrid system that integrates learned task controllers and a neural 3D safety monitor into conventional industrial workflows to automatically perform flexible cable insertion and soldering on a motor production line. Using less than 20 minutes of real-world data per task, the system ran continuously for 5 hours 10 minutes without physical fencing, producing 108 motors with a 99.4% pass rate on product-level quality inspection, a takt time close to that of human workers with smaller cycle-time variation, and solder-joint quality comparable to that of skilled operators.
Why it matters: The paper reports data from continuous operation of learning-augmented industrial automation on a real production line for 5 hours 10 minutes, with 108 motors and a 99.4% pass rate, serving as a practical reference for learned control entering manufacturing.
The authors propose an adaptive collision sensitivity framework that dynamically estimates collision force based on the effective mass (EM) of each robot link and uses it to decide whether to interrupt motion, instead of applying a fixed contact threshold to the whole robot. Experiments were conducted on a real UR10e robot with AIRSKIN sensitive skin and in simulation, supplemented by simulated experiments on a KUKA LBR iiwa using joint torque sensing. In a mock pick-and-place scenario including both transient and quasi-static collisions, production efficiency improved by more than 45% compared with the standard method.
The paper proposes agentic manipulation policies (AMP), in which a generalist AI model directly controls a robotic arm and adjusts its actions as the task progresses. The authors' reference implementation, AMP1, runs unchanged across single-arm, bimanual, and mobile manipulation.
Why it matters: The paper reports zero-shot success rates and real-world trial data for a generalist AI model directly controlling a robotic arm, which can be compared with specialist policies such as VLA and WAM.
SURE is a data-efficient vision-language-action (VLA) training framework that models spatial uncertainty with a lightweight transposed-convolutional probabilistic action decoder and uses Soft-DTW alignment supervision to accommodate temporal variation across demonstrations. On LIBERO, RoboTwin 2.0, and real manipulation tasks, SURE matches or exceeds baseline VLA manipulation success rates using only 10% of demonstration data or 20% of training steps.
ODEAction proposes a chunk-free action generation paradigm that replaces the action-chunk prediction common in VLA models with continuous action evolution. It characterizes the rate of change of the current action over evolution time using a transition field over action space; during execution, it takes the current action as the initial condition and recursively generates subsequent actions via ODE integration. The method is implemented as an action head and does not depend on a specific vision-language backbone architecture. Experiments on LIBERO, EBench, RoboTwin, and real robots show better task performance and trajectory quality than matched action-chunking baselines, with particularly notable gains on fine manipulation tasks.
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).
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 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 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 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.
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.
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.
The research team proposes A2-MEAS, an embodied multi-agent framework that uses code as a scaling interface. It combines correction codes with corrective language, spatial evidence, execution history, and action proposals, while MetaBrain provides training 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 presents comparison data across multiple benchmarks and real-robot experiments.
Sun Yat-sen University, Peng Cheng Laboratory, and other institutions propose RePO-VLA, which segments recovery trajectories, trains a progress-aware semantic value function, and performs value-conditioned fine-tuning so that dual-arm VLA policies can continue advancing tasks after disturbances. The study also builds FRBench, a failure-recovery benchmark; the RoboTwin simulation contains 46 dual-arm tasks and 23,453 trajectories. On real robots in water-pouring and towel-folding tasks, as recovery demonstrations per task increase from 50 to 200, the average disturbed success rates are 30%, 65%, and 75%, respectively.
Tristan, co-founder of YC F26-incubated company Asim, said the team is building a standardized benchmark for physical intelligence in humanoid robots, with a focus on evaluating dexterous manipulation capabilities. The benchmark uses repeatable scenarios to examine spatial reasoning, grasp adaptation, coordination, and how robots respond to physical interaction, with the goal of pinpointing where capabilities fail and providing reproducible failure cases. The team also noted the need to address comparability issues caused by differences in hardware, sensors, and control interfaces.
The authors propose Warp, a context-conditioned monotone bijective spline adapter that nonlinearly reshapes the sampled outputs of a frozen base policy while preserving per-dimension ordering. Across four ManiSkill tasks, Warp achieves the strongest recovery in a matched evolution-strategy adapter study; after switching to PPO, the ordering relationships vary by task. In a physical SO101 peg insertion task, Warp's success rate is comparable to a strong stochastic residual, with median time per successful cycle reduced by about 30%.
RoboTryOn proposes treating human-to-robot translation as virtual try-on: it preserves the demonstrated motion, objects, and background, and only replaces the human with the target robot specified by a reference image. The method autoregressively converts keyframes and uses them to guide video synthesis, with motion conditioned on the human hand skeleton; it then uses an inverse dynamics model (IDM) with depth features to infer action labels from the synthesized videos for policy learning. Experiments show that policies trained only on synthetic data improve average success rate from 2% to 60% on new tasks, including dual-arm manipulation, and reach the level of policies taught by pure teleoperation at about one-third of the collection cost.
This paper investigates how inertia mismatch between a robotic arm and the objects it manipulates affects robot policy learning. Through controlled simulation experiments that vary robot and object masses, it finds that reinforcement learning succeeds across a broad range of mass combinations, while behavior cloning learns more reliable policies when a lighter robot manipulates a heavier object.
The authors propose Tactus, a textile-integrated sensing glove that provides both capacitive tactile force sensing and IMU-fused pose tracking with a 1.5 mm fingertip stackup thickness, for collecting natural dexterous manipulation demonstration data. At 95% confidence, the glove's minimum detectable force is 0.59 N, 1.6–2.7 times lower than two commercial tactile glove baselines, with cross-subject consistency of 0.87; under the most severe manipulation occlusion, the relative error of fingertip keypoints is 11.0 mm.
The roughly $200 Waveshare RoArm-M2-S four-degree-of-freedom robotic arm, under a fixed overhead camera, uses a YOLO11n-OBB detector, an ArUco reference pose, and a color classifier to perform color sorting of writing utensils without a wrist rotation joint. For utensils deviating more than 45 degrees from the fixed approach direction, corrective sweeps bring them into a graspable pose within at most 4 attempts. In 196 single-utensil trials, 176 (89.8%) landed in the correct color bin; of the 20 failures, 19 were grasp slips or missed grasps and 1 was a color mislabel.
This paper proposes a teleoperation-cost-effective framework that uses temporally reversed trajectories from easy tasks as supervision signals for hard manipulation tasks, reducing manual demonstration collection. The framework has three parts: a closed-loop data collection pipeline, hierarchical data refinement based on kinematic priors and critic-guided advantage filtering, and iterative policy learning. In two simulation benchmarks and real-robot experiments, the method improved success rates on hard tasks compared with reversal-based and reinforcement-learning baselines, with higher data efficiency and more stable training.
The authors propose Selected Diffusion Noise (SDN), a training-free method that treats diffusion noise as a controllable degree of freedom to improve the test-time reliability of VLA policies. SDN samples multiple candidate action chunks and performs two-stage selection: it first filters by task-relevant visual grounding using a counterfactual negative set, then selects the kinematically most stable action chunk based on JerkRMS. In simulation and real-world experiments with three VLA policies, it achieved improvements of 8-10% and 18.33%, respectively.
The study proposes the InDex framework, which uses a normalized aperture signal derived from retargeted demonstrations to supervise a scalar grasp intent, adapting parallel-gripper VLA models to multi-fingered dexterous hands in two stages. Across four simulated tasks, InDex with π0.5 achieved 85.8% mean task success, a 35.5 percentage point improvement over the native policy, and reached 69.0% on new objects with similar geometry without fine-tuning. It was also effective across three VLA backbones and on a separately trained real-world policy, whose mean success rate rose from 71.0% to 83.0%.
This study compared retry mechanisms of code-as-policy agents on simulated robot manipulation tasks: on the same 252 initial failure cases, Qwen2.5-Coder-7B, using feedback templates selected by failure labels, recovered 85.3% within five retries, versus 68.3% for resampling with the same retry cap, and the advantage persisted under a fixed total token budget.