A study published in the International Journal of Automation and Control proposes a genetic algorithm-optimized fuzzy controller (GA-FLC) that reduced average root mean square error by 89% compared with conventional feedback control in simulations of a three-degree-of-freedom robotic arm.
Tsinghua University and UC Berkeley jointly proposed and open-sourced uni-robot-agent, a cross-environment unified embodied agent harness for connecting tool interfaces across different robots and simulation environments. Based on this framework, Qwen3.8-Flash-Next scored 31.39 on the RoboDojo benchmark, exceeding the publicly reported 28.97 for GPT-6 Astra, at about 1/20 of Astra's cost.
NTU MMLab and Shanghai AI Lab propose PhysLDM, which uses a spatiotemporal VAE to compress an entire deformable-object simulation trajectory into latent space and then generates it in one pass with a diffusion model, achieving an average reconstruction error of 2.48 millimeters.
The paper LT-Mem from the APRL lab at Daegu Gyeongbuk Institute of Science and Technology (DGIST) in South Korea won the IROS 2026 Best Paper award. The work proposes a volatile-perception spatiotemporal memory framework that uses three memory layers—Live, Delta, and Meta—and three volatility-based update strategies—coverage, retention, and multi-hypothesis—to solve the temporal forgetting problem in which new observations overwrite old memories when a robot revisits the same space multiple times.
An ETH Zurich team has developed FlexiQuad, a ring-shaped flexible-frame quadrotor with no central body; on impact the frame deforms to absorb energy and the drone can keep flying. It reaches a top flight speed of 80 km/h, with collision tests at about 25 km/h; the frame material costs less than 15 Swiss francs, and it has been used to grasp and carry soft first-aid kits and payloads heavier than its own weight. The team also used reinforcement learning to let it stabilize its attitude after a collision before resuming autonomous flight, but reliable perception and autonomous navigation remain prerequisites for leaving the lab.
Zou Danping's team at Shanghai Jiao Tong University proposes DiffDreamer, which couples a learned world model with differentiable simulation for agile quadrotor control; the paper has been accepted to RAL 2026. The method uses the world model to predict external dynamics and differentiable simulation to propagate the robot's physical state. In aggressive tracking tasks, distance error drops to 0.381 m, versus 0.793 m for the strongest baseline; in dynamic gate crossing at gate speeds of 1.0/1.5/2.0 m/s, the number of gates passed is 16/12/8, versus 12/4/1 for MPC.
The Xbotics Embodied Intelligence Lab built a closed-loop simulation pipeline using the AgiBot GE-Sim 2.0 world model and the OpenPI π0.5 policy. In testing, the world model generated about 4.33 frames per second, the pure-policy closed loop ran at about 3.20 frames per second, and after adding MiMo online scoring, the rate dropped to about 0.88 frames per second.
Why it matters: The author uses a four-stage test to break down the frame-rate bottlenecks and causes of reward hacking in world-model closed-loop simulation, which can be transferred to the design of VLA evaluation workflows.
XPENG ROBOTICS proposes the embodied vision-language model Capek 0.5, training separate experts for four execution capability categories—spatial reasoning, temporal understanding, action guidance, and state verification—then integrating them into a single model via TIES merging and routed MOPD distillation.
Lingchu Intelligence, together with Beihang University, Peking University, and The Chinese University of Hong Kong, proposed a frequency-aware flow matching method called FAFM; the paper was accepted to the NeurIPS 2026 main conference. The method uses the discrete cosine transform (DCT-II) to represent actions as frequency-domain coefficients corresponding to real time, performs flow matching in coefficient space and constrains the first-order time derivative, requires no additional learnable parameters, and can be used independently or integrated into VLA models such as π₀ and π₀.₅.
Why it matters: The paper presents a frequency-domain action representation that adds no learnable parameters and can be integrated into existing VLA models, offering a transferable approach for joint training on heterogeneous-frequency data.
RoboFFT, proposed by the PKU-Spirit AI Joint Lab, has been accepted to CoRL2026. It is a reinforcement learning post-training method for generative robotic policies that reuses the forward noising process from diffusion model and flow model pretraining to construct an approximate probability ratio for PPO-style clipped updates.
Why it matters: The method reuses the forward noising loss of diffusion and flow models to construct policy update signals, offering a reusable approach for post-training generative policies.
Whether you're doing BC or RL, RL can help collect high-quality data. It can also stack two cubes on top of another cube. A fun paper with @qiyang_li @zhiyuan_zhou_ @CharlesXu0124 on how RL assists robot data collection!
KEC combines Fable with Astra to rapidly tune complex simulation parameters via RSS, including geometry, dynamics, deformation, materials, lighting, textures, and camera extrinsics, so that the simulation matches a given input video. This work targets the capability gap in agentic robotics for extending to new environments, new objects, and new tasks.
G1 Brunson uses event-based rewards and asymmetric actor-critic training, bringing the depth camera directly into the observation loop during training to form a G1 whole-body dribbling policy. Reproducing the official weights locally in MuJoCo gave an average of 43.5 crossovers over 12 episodes with zero falls, matching the official results.
WAM trains a forward-dynamics world model on quadruped robot locomotion data collected under convex model predictive control, replaces hand-crafted steering laws with receding-horizon sampling planning, and combines A* global paths with dynamic obstacle extrapolation to achieve navigation and obstacle avoidance. The material states that all numerical values come from local reproduction experiments.
The study proposes H2RB, a human-robot co-training framework that adapts pretrained VLA policies on unpaired human and robot demonstrations through geometry-anchored alignment and a decoupled-decoder tokenizer. In real-world tasks, object transfer performance improved from 8.0 to 51.8, and the skill transfer score reached 62.7, while the co-training baseline completely failed to transfer human demonstration skills. Out-of-distribution transfer under object placement changes also improved from 6.0 to 45.3.
A research team at Seoul National University has proposed RISE (Responsibility-Induced Specialized Experts), a VLA framework that can collect demonstration data for locomotion and object manipulation separately and then learn jointly from them, enabling a robot to execute multi-stage tasks consecutively without full multi-stage demonstrations.
The team led by Chunhua Shen at Zhejiang University, together with the China Telecom Artificial Intelligence Research Institute (TeleAI) and Chenjia Bai's team at Gamma Intelligence, proposed GR-BC, which builds multiple attempts at the same task into a graph, uses reverse Dijkstra search to estimate the remaining steps from each state to success, and then weights action chunks accordingly to train a vision-language-action model.
Nanyang Technological University and other institutions present MemSkill at NeurIPS 2026, packaging memory operations as learnable, reusable, and evolvable Memory Skills, with a Controller selecting skills, an Executor building memories, and task feedback used to optimize the selection strategy.
AMI, the world model company founded by LeCun, together with New York University, INRIA Paris, and Brown University, released a paper on H-JEPA, proposing a hierarchical world model that stacks multiple JEPAs layer by layer for visual planning.
A research team from Nanyang Technological University and other institutions presented MemSkill at NeurIPS 2026, packaging memory operations into learnable, reusable, and evolvable Memory Skills, optimizing skill selection through task feedback, and revising and expanding the skill library based on failure cases.
Wang Rui, a 2022 undergraduate student in the School of Future Science and Engineering at Soochow University, published research on the TACHIP tactile chip in Nature Communications as a co-first author under the guidance of Professor Chen Tao's team. The plug-and-play soft tactile chip uses a five-pixel pneumatic array to reconstruct tactile patterns such as shape, hardness, roughness, and temperature, achieving a 95% first-round recognition rate and raising the success rate of six types of teleoperation tasks from 65% to 85% through reverse tactile feedback. The team completed validation demonstrations on the EFORT industrial robot platform.
A University of Bristol team built a tabletop physical model simulating muscle contraction using a row of ordinary electric motors coupled to each other through a shared mechanical skeleton, with results published in the Royal Society journal Interface. In the experiments, the motors made intermittent contact with the skeleton to transmit force, and without direct communication, muscle-like behaviors spontaneously emerged, including a load-velocity relationship, more units participating as load increased, and coordination waves. The researchers believe this approach, in which coordination is handled by the structure itself, could help improve how robots are actuated.
Brain-computer interface company Junrou Technology has completed a series of angel round financings: the angel round was led by Jiuhe Venture Capital, and the angel+ round was invested in by Shunwei Capital, Hillhouse Venture Capital, Agibot, and others, with the three rounds completed within one year of the company's founding totaling 100 million yuan. The company's core technology is a neural worm dynamic electrode originating from Liu Zhiyuan's team at the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences; a planar electrode array is rolled into a fiber electrode about 196 micrometers in diameter, which crawls toward multiple brain regions under magnetic field guidance through a small hole of about 1 centimeter in the skull. The first-generation brain-computer interface system has completed some animal experiments, and an IIT is expected to launch in 2027.
The SoftCon framework builds an octopus swimming control pipeline from tetrahedral soft-body dynamics, muscle constraints that contract along fibers, and traveling activation waves generated by keyframe matrices and propagation filtering, using PPO to learn coordination rather than reconstruction. The framework originates from SIGGRAPH Asia 2019.
G1 Dribble uses end-to-end reinforcement learning to achieve humanoid robot soccer dribbling; the walking task's weights are not adjusted, and velocity commands drive both the body and the ball. The author tested it locally in MuJoCo-Warp and recorded eighteen cloud sessions, reporting touch rhythm, the distance band for re-touching the ball, and failure modes, including that a centered start steps over the ball without touching it, a first touch occurs only when offset to the side, and at 1.5 meters per second it knocks the ball away directly.
HACMan++ targets long-horizon manipulation for six-dimensional object rearrangement, decomposing actions into three parts: selecting a primitive, anchoring a placement point on the point cloud, and continuous parameters, with unified decision-making via a per-point, per-primitive evaluation map. On this basis, the framework achieves closed-loop chaining of pushing, grasping, moving, and placing.
Neural-Fly proposes a residual aerodynamic force adaptation law that combines wind-invariant representations with online least squares for online adaptive control of quadrotors in strong winds. The method trains a network on public flight data and conducts comparative figure-8 trajectory tracking experiments with a PyBullet quadrotor in a steady wind field, demonstrating rapid adaptation enabled by domain-shift separation.
South Dakota State University and other institutions propose Any3R, which uses a dense ray field to describe arbitrary central cameras, reconstructing a 3D point cloud and estimating pose from mixed-camera images in a single forward pass. The method replaces the row and column indices of rotary position encoding with angular coordinates and weights the pixel loss by solid angle, so one set of weights covers pinhole, fisheye beyond 180°, and 360° panoramic cameras. On benchmarks spanning 13 central camera types and 3 datasets, it leads in both reconstruction accuracy and pose accuracy; the inference code, 13 camera models, and weights are open-sourced.
Why it matters: The article presents a feed-forward 3D reconstruction method that replaces the pinhole assumption with a dense ray field, offering insight into handling mixed inputs from multiple camera types.
A team led by Cheng Long at the Institute of Automation, Chinese Academy of Sciences, published research at IROS 2026 using tactile-enhanced reinforcement learning to complete sub-millimeter precision insertion, raising the success rate from 52% for a tactile-free approach to 67% under a 0.05 mm clearance.
GenSight is a unified embodied world model that integrates future video generation and dense perception into a single pretrained video diffusion Transformer. It uses a unified spatiotemporal representation covering video, depth, embodiment segmentation, and 2D optical flow relative to the first frame; built on a shared DiT backbone, it predicts each modality within the same architecture via lightweight task modulation, and jointly optimizes them in a unified velocity prediction space. At inference, it supports multi-step future video generation with aligned dense prediction, as well as single-step dense perception from an observed video. Training data covers a variety of robot embodiments and human hand interaction.
The paper proposes a Transformer-based architecture with shared modular recurrence for universal morphology robot control and evaluates its generalization performance on a set of MuJoCo robots. The authors note that contextual features are often only partially available but can be recovered through modular interactions, improving multi-robot control and generalization to contexts unseen during training. Results show substantial improvement in zero-shot generalization performance on robots with unseen dynamics, kinematics, and topologies across four environments.
The authors present MUSBRIC, a modular hardware platform and accompanying MuJoCo simulator for reconfigurable musculoskeletal robots, which uses LEGO-compatible endcaps to turn fiber-reinforced extensile pneumatic artificial muscles into reusable, plug-and-play muscle modules. An elbow-joint prototype validated load-dependent flexion and extension, with pulley-guided routing enabling reversible motion and a mechanical stop stabilizing the straight posture; a robotic hand demonstrated compliant grasping and multi-joint construction, 15 leg configurations completed matched sweeps in hardware and simulation, and a quadruped assembled from four legs achieved walking. The hardware design and simulation code will be open-sourced.
LOGIV is a task-graph framework for long-horizon robot tasks that connects symbolic planning, observation-based verification, and embodied execution. During execution it checks the physical validity of condition-support and resource-ordering edges, and after a failure it uses dependency tracing to determine which parts must be restored.
OpenFlyV2 is a unified onboard vision-language-action architecture that uses a shared vision-language representation to support semantic understanding, language-grounded perception, spatial reasoning, and continuous flight control within a single model. The authors built a simulation dataset of approximately 70,000 navigation trajectories and 60,000 vision-language instruction-response pairs. On an NVIDIA Jetson Orin NX, the model runs fully onboard and generates action chunks of 12 waypoints each at 10 Hz.
The paper proposes actuator-aware (AA) observer inputs: the altitude and attitude observers use limited virtual inputs filtered by an assumed rotor model and aligned with the measurement delay, while the horizontal observers use acceleration reconstructed from measured attitude.
SimRetract proposes a policy trained only on simulated demonstrations that lets a robotic arm take a retractor from a surgeon's hand and maintain retraction at handoff, achieving zero-shot sim-to-real transfer for surgeon-robot collaboration in open surgery.
The research team released TacClaw-Infinity, a large-scale structured Vision-Tactile-Language-Action (VTLA) dataset for contact-rich manipulation, spanning over 10,000 hours, 1 million episodes, and 1 billion frames across more than 1,000 task types.
Why it matters: This dataset provides a collection interface and quality control pipeline for ten-thousand-hour-scale tactile manipulation data, and its cross-embodiment transfer validation approach can serve as a reference.
The research team proposes I3L, a hardware and software system that combines whole-body correction with iterative policy refinement to mitigate the out-of-distribution degradation of offline imitation learning policies in long-horizon whole-body mobile manipulation.
The authors use a sparse autoencoder (SAE) to analyze the internal representations of a VLA model, identify a feature whose activation strongly correlates with the presence of adversarial patches, and suppress that feature at inference time only when a linear probe detects an attack, thereby improving robustness without fine-tuning the VLA. In adversarial patch attack tests on LIBERO-10, conditional intervention improved the success rate under intermittent attacks, whereas continuously applying the same intervention substantially degraded policy performance.
Based on the dynamics-aware diffusion-denoising framework D4orm, this paper builds a family of multi-robot trajectory planners for different operational requirements, including a decoupled planner, an online receding-horizon planner with feedback control, and a distributed planner. Evaluations with differential-drive and holonomic robots in 2D and 3D environments show that D4orm-based methods find high-quality solutions faster and more reliably than sampling-based optimization methods such as MPPI and a learning-based diffusion model method.