Content types
Research papers
Papers, technical reports, research datasets, and new benchmark designs centered on a new method, experiment, or finding. Releasing code or a physical robot demonstration alongside a paper does not change its main research contribution.
105 selected · 72 in the last 30 days
机器人前瞻 · 微信公众号 ·
NVIDIA open-sources robot data engine Flora, labeling cost about 1/125 that of GPT-6 Astra
NVIDIA released and open-sourced the data engine Flora, which converts raw robot video into training data such as timestamped action descriptions, process rewards, and outcome rewards; it is already used to train the GR00T series models. Flora consists of an action description model, a temporal segmentation model, and a reward model, fine-tuned from the Qwen3.5 series of vision-language models, and can label 200,000 hours of data per day on 512 GB300 GPUs.
arXiv cs.RO ·
Activation injection study dissects vision-language-action models
Researchers propose an interpretability pipeline to analyze what a fine-tuned vision-language-action (VLA) policy is computing internally. On pi0.5, injecting task A's activations into the action expert while the policy views task B's scene caused 99.6% of episodes (n=1968) to execute task A's motor trajectory, indicating the injected program is bound to absolute workspace coordinates and ignores the visible scene.
arXiv cs.RO ·
SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model
SuperMimic uses a differentiable world model of shared dynamics to train the world model and the tracking policy together in a single training loop, giving the policy analytic gradients through backpropagation-through-time over short imagined rollouts.
Xbotics具身智能实验室 · 微信公众号 ·
SenseTime launches SenseNova-RoboRSI and a 27B embodied model, achieving SOTA on 12 simulation benchmarks
The SenseTime SenseNova team launched the SenseNova-RoboRSI unified execution framework and the 27B embodied model SenseNova-RoboRSI-27B-Preview.
智源研究院 · 微信公众号 ·
BAAI's FlagOS 2.2 released: over 1,454 operators, 27% cross-chip performance gain
BAAI presented FlagOS 2.2 progress at the 2026 AI Open Computing Conference and FlagOS Technical Conference; the version was released on September 30. Public operators in the nine operator libraries rose from 564 to 1,454 (890 added). In tests covering over a thousand operators and 10 heterogeneous AI chips, cross-chip operator performance improved by about 27% over the previous version. The FlagOS-Compressor weight conversion and quantization tool was released for the first time, enabling a new model to be adapted to as many as 12 chips within one day.
自变量机器人 · 微信公众号 ·
X Square Robot releases X2Real simulation benchmark, with 0.96 correlation between simulated and real-robot task success rates
X Square Robot has released X2Real, an evolvable simulation evaluation benchmark. Its simulated evaluation and real-robot evaluation achieve a task success rate correlation of 0.96 and a progress score correlation of 0.99.
arXiv cs.RO ·
Evaluating a Frozen Frontier VLM Directly as a Robot Policy on the RoboMME Memory-Dependent Manipulation Benchmark
The study uses a frozen frontier vision-language model, GPT-6-Astra, directly as the robot policy and tests it on RoboMME, a benchmark of 16 memory-dependent manipulation tasks. The model keeps the entire episode context throughout and directly outputs each end-effector waypoint and gripper command, with no training specific to this benchmark.
Dexmate ·
Dexmate and CMU Release MeMoREval Benchmark: Robot Storing Camera Frames Beats Text Descriptions
Dexmate and CMU introduced the MeMoREval benchmark, which uses a two-stage protocol of experiencing first and asking questions later to evaluate what embodied agents should store as memory. It includes 53 tasks and 4 simulated indoor environments, and runs on the Vega 1 Pro dual-arm mobile manipulator. Results show that storing raw camera frames scored 70% and had lower write cost, surpassing text descriptions at 56 to 60%; adding scene graphs achieved the highest overall success rate among all tested models. The protocol also completed two memory retrieval demonstrations on a real Vega 1 Pro using a SLAM map and head camera.
CMU Robotics Institute ·
CMU research team develops human-inspired multisensory fingertips for DeltaHands robotic hand
CMU researchers designed human-hand-inspired fingertips that can be mounted on the DeltaHands flexible robotic hand. They combine strain gauges to measure pinch pressure with a microphone that captures micro-vibrations through a plastic fingernail, enabling the hand to count paper cups, identify materials, and grasp fragile items such as tofu cubes and potato chips. Each fingertip costs less than 100 US dollars, compared with about 800 US dollars to build a similar setup from off-the-shelf products. The work was presented at the 2026 Intelligent Robots and Systems Conference in Pittsburgh.
Sanctuary AI ·
Sanctuary AI releases TIGER method to accelerate robot reinforcement learning with imitation policies as dense rewards
Sanctuary AI proposes TIGER (Task-Space Imitation Guidance for Efficient Reinforcement Learning), which does not execute the imitation policy but instead maps its action chunks through the controller into end-effector reference trajectories, converting them into dense rewards for reinforcement learning.
Singularity Hub ·
KAIST quadruped robot RAIBO2 completes marathon without battery swap, ~40 miles per charge
KAIST's quadruped robot RAIBO2 finished the Persimmon Marathon in Sangju, South Korea, in 4 hours 19 minutes 52 seconds, slightly faster than the human average, without swapping or recharging its battery, and still had 34% charge at the finish; the team described the work in a Nature paper last week.
CMU Robotics Institute ·
CMU and Siemens Partner to Propose PRISM, a Safety-Stop Monitoring Method for Humanoid Robots
CMU and Siemens partnered to develop a method for evaluating the "safe stoppability" of humanoid robots, and used simulation to define the conditions under which a robot can be safely stopped. The result is PRISM (Proactive Refinement of Importance-sampled Stoppability Monitor).
Skild AI ·
Skild AI releases physical self-play post-training results; S1 model learns soccer in simulation and transfers to humanoid robots
Skild AI announced that it post-trained its foundation model S1 through physical self-play, letting the policy play against a recent version of itself with scoring as its sole objective. In NVIDIA Isaac Sim, over 140 years of simulated time, it learned skills such as dribbling past opponents, shielding the ball, and tackling, and these skills were successfully transferred to humanoid robots for matches. The authors believe self-play can break through the capability ceiling of pretrained human data, and the next step will be to extend it to everyday tasks such as collaborative manipulation and city-scale navigation.
MagicLab/魔法原子 ·
MagicLab analyzes Magic-Brain's in-context visual guidance design; magicVLA grasp success rate rises from 70% to 94.4%
MagicLab introduces the visual guidance approach in the Magic-Brain framework: a VLM Agent outputs a bbox, and a segmentation model refines it into a three-color mask overlaid on the observation image, converting semantic guidance into visual guidance.
PKU EPIC Lab ·
PKU EPIC Lab proposes UniPart for zero-shot language-driven 3D part segmentation
PKU EPIC Lab proposes UniPart, a feed-forward cross-modal 3D Transformer conditioned on CLIP text embeddings that uses free-text phrases to select functional parts on point clouds, achieving zero-shot language-driven 3D part segmentation.
PKU EPIC Lab ·
FoldNet++: A Large-Scale Synthetic Dataset for Robotic T-Shirt Folding and Unfolding
PKU EPIC Lab has released FoldNet++, a large-scale synthetic dataset for robotic T-shirt folding and unfolding, covering 6 robotic embodiments, 1K T-shirts, 1K environmental assets, and 120K richly annotated demonstrations.
MIT CSAIL ·
MIT CSAIL and others launch InstructMesh tool to repair AI-generated 3D models and manufacture them on demand
MIT CSAIL, together with Google and Northeastern researchers, launched InstructMesh, which combines TRELLIS with GPT-4 to let users repair AI-generated 3D models using natural language before 3D printing them.
PKU EPIC Lab ·
GIF: An agentic generation framework for interactive and functional object compositions for robot learning
PKU EPIC Lab proposes GIF, an agentic generation framework for robot learning that reframes the problem as disentangled reconstruction followed by relative pose recovery. CoGen combines 2D and 3D generative models to produce instance-disentangled meshes and coarse initial poses; GPRM refines the relative poses under joint geometric and physical guidance; and a VLM verifier selects the candidate that best matches the structured specification.
中科慧灵 · 微信公众号 ·
CASIA and others propose World4RL: diffusion world model combined with reinforcement learning to optimize robot manipulation policies, real-world success rate rises to 93.3%
The Institute of Automation of the Chinese Academy of Sciences, Zhongguancun Academy, and the CASBOT team propose World4RL, a policy optimization framework for robot manipulation that combines a diffusion world model with reinforcement learning, enabling robot policies to be optimized end-to-end in imagined trajectories generated by the model without additional real-robot interaction.
PKU EPIC Lab ·
ZETA: A Controlled Study of Zero-Shot Cross-Embodiment VLA Transfer for Tabletop Manipulation
PKU EPIC Lab published the ZETA study on arXiv (arXiv:2609.02546), proposing a unified definition of zero-shot cross-embodiment VLA transfer and a controlled benchmark covering 14 held-out target embodiments, distinguishing strict zero-shot transfer from pretrain-exposed transfer. Experiments show that local end-effector state-action representations, source embodiment diversity, and auxiliary co-training improve transfer by about 15, 18, and 7 percentage points, respectively, and that adding only 5% target-embodiment data during pretraining raises average target-embodiment progress by 13.4 percentage points.