Skip to content
#Robot models and learning
Oct 19Mon
  1. OFweek Robotics62

    OpenAI announces in-house humanoid robot; author interprets its motives and impact on the industry

    Altman said OpenAI will develop its own humanoid robot, with a demo in 2027. The near-term goal is to assist skilled workers in building infrastructure, while the long-term goal is a personal robot. The author reviews OpenAI's two previous retreats: the 2019 robotic hand that solved a Rubik's Cube, the disbanding of its robotics team in 2020, and Figure's announcement in February 2025 that the partnership was ending. The author believes OpenAI now has the necessary conditions, including Aditya Ramesh's world simulation project, Astra's virtual capabilities, simulation infrastructure, and a teleoperation team of more than 100 people in San Francisco.

  2. 中控技术38

    SUPCON Releases TPT Industrial AI Foundation Platform and AOP Autonomous Operations Factory System, Holds Industrial AI Results Conference in Ningbo

    SUPCON held an industrial AI results conference in Ningbo, launching an industrial AI foundation platform centered on the TPT time-series large model and integrating it with the universal control system UCS to build the AOP autonomous operations factory system, connecting the perception-analysis-decision-execution chain. TPT and AOP have been replicated across 29 sub-sectors, and SUPCON signed industrial chain cooperation agreements with Huawei, Alibaba Cloud and others, while the industrial AI data alliance has more than 300 members. Companies including Wanhua Chemical and Zhenhai Refining & Chemical disclosed the results of TPT in early warning and optimized operations at chlor-alkali and refining units.

  4. 深蓝AI · 微信公众号62

    Shenlan College livestream preview: Robot motion control from two-stage to one-stage

    Shenlan College will hold a livestream on October 21 titled "Robot Motion Control from Two-Stage to One-Stage," featuring Ye Lei, a PhD student jointly trained by Harbin Institute of Technology and Shanghai Innovation Institute. Centered on three works—KCFRC, PegasusFlow, and PredActor—the session will explain how robot motion control moves from a two-stage approach of planning first and then tracking to a one-stage approach that jointly generates states and actions.

  7. arXiv cs.RO67

    Metro-WM: Long-horizon latent planning with realizable sub-goals

    Metro-WM proposes a hierarchical planning framework that, instead of generating unconstrained latent sub-goals, retrieves real states from observed frames in offline expert demonstrations or random-action trajectories as sub-goals, and connects frames from different episodes into paths toward the goal. Experiments show its long-horizon success rate is up to 37.33 percentage points higher than the next-best hierarchical method, up to 10.9 times faster, uses 13-56 times less offline computation, and requires less tuning. The authors also note that leading macro planners frequently output physically unrealizable sub-goals.

  8. arXiv cs.RO57

    EgoTac: Predicting In-the-Wild Touch from Egocentric Vision

    EgoTac is a generalizable model that predicts rich tactile information directly from egocentric human videos. It was trained on more than 5.7 million image-tactile pairs covering both continuous force measurements and binary contact. Experiments show an average force error below 0.06N for in-domain prediction, and on out-of-domain contact prediction benchmarks it consistently outperforms the current state-of-the-art contact estimator. It also enables zero-shot prediction on unconstrained real-world videos; scaling analysis indicates that both data diversity and data volume steadily improve performance. The work was accepted to NeurIPS 2026.

  10. arXiv cs.RO70

    WorldOracle: Benchmarking action-conditioned world models against real-world executions

    WorldOracle is a benchmark for action-conditioned world models (ACWMs) that uses identical initial observations and low-level action sequences to test whether predicted rollouts can reproduce real-world physical consequences and task-state transitions. The evaluation covers four dimensions: physical state correctness, interaction spatial alignment, key interaction area fidelity, and multi-view action coherence, and is paired with long, multi-segment real-world data containing both successful and failed executions. Experiments on seven representative ACWMs reveal different models' strengths and failure modes. The dataset, evaluation code, and leaderboard are publicly available.

  11. arXiv cs.RO78

    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.

    Why it matters: The paper provides complete comparative data on a frozen frontier VLM used directly as a policy on a memory-dependent manipulation benchmark, which can be used to assess how much of the embodied stack a general-purpose model can take over.

  12. arXiv cs.RO47

    Evaluating Sensitivity of World-Action Models to Spatial Instruction Rewriting

    An evaluation conducted 864 rollouts of three world-action models across four RoboLab scenes, holding initial states and task goals fixed, and found that describing spatial relations using the reference object rather than the object to be moved lowered the stable placement success rate by an average of 12.2 percentage points. An independent test of the policies' language understanding identified relation reversals, and adding images disrupted otherwise correct text-only interpretations. Rollout analysis also distinguished spatial placement from instruction following: the stable placement rate fell from 43.2% at any time to 21.8% at episode end.

  13. arXiv cs.RO56

    Analytic Robot Self-Model: Learning Screw-Axis Kinematics for Rendering and Model-Based Control

    The paper proposes an explicit analytic robot self-model that requires neither deformation networks nor URDF/CAD templates. From multi-view images and joint readings, it obtains appearance and rigid link assignments, and independently identifies metric screw axes using sparse end-effector poses, yielding closed-form forward kinematics and an analytic Jacobian that can be extrapolated.

  14. arXiv cs.RO67

    RoboHarness: Simple Robot Harness Lets LLM Agents Outperform VLA and World Action Models on BEHAVIOR Challenge 2025

    An arXiv paper presents RoboHarness, a robot harness that gives LLM agents a visual-geometric control panel so they can directly understand and invoke embodied tasks. The authors argue that coding agents have the potential to achieve generalist embodied manipulation without the route of large-scale VLA training. Experiments show that, with RoboHarness, Qwen3.8-Flash-Next outperforms state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks, and the implementation is open source.

  15. arXiv cs.RO45

    Reinforcement Learning-Based Tendon-Driven Bionic Jellyfish Robot Achieves Closed-Loop Depth Control

    A tendon-driven bionic jellyfish robot achieves 150° bending through soft actuators that combine a flexible substrate with discrete constraints, with tendon displacement and bending angle approximately linear. Eight actuators driven by four servos enable stable swimming, attitude adjustment, and self-righting. A reinforcement learning controller trained on this linear actuation characteristic achieved closed-loop depth regulation in both simulation and physical experiments.

  17. arXiv cs.RO60

    Improving legged locomotion policy health via gait-phase-conditioned effective rank of the network Jacobian

    This study analyzes reinforcement learning policies for legged locomotion using the effective rank of the network Jacobian, finding that conditioning on gait phase exposes representation structures invisible from the global rank. Architectures with observation normalization and residual connections allocate more than one additional effective-rank dimension to swing-dominant states than vanilla MLPs do. Based on this, the proposed diagnostic method reduces joint jitter by up to about 2x in simulation and on a physical Spot, and the paper has been accepted to CoRL 2026.

  18. arXiv cs.RO72

    ADEPT: Accelerating Dexterous Manipulation via Reinforcement Learning Pre-Training and Post-Training

    ADEPT is a large-scale reinforcement learning framework that first pre-trains a dexterous policy on a generic object-reposing task, then post-trains downstream policies with this prior, enabling long-horizon tasks to be solved directly from raw visual-tactile perception across high-DoF embodiments. To prevent naive fine-tuning from degrading zero-shot reposing ability, the authors propose a stable post-training scheme combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates, and introduce a joint-space Geometric Fabric that coordinates the policy with the robot.

  19. arXiv cs.RO57

    Cyclops: Synthesizing Sparse LiDAR Intensity into RGB Video for Camera-Free Perception

    Cyclops is a framework that synthesizes sparse Non-Repetitive Scanning LiDAR (NRS-LiDAR) intensity data into RGB video for all-weather perception tasks without a camera. Its method first uses a frozen pre-trained densification module to convert sparse intensity projections into dense representations, then learns a velocity field through Latent Bridge Matching (LBM) to transport intensity latents to the RGB distribution within a few ODE integration steps; it also injects prior-frame context via temporal attention to suppress inter-frame flickering, and models the velocity field as a policy optimized by a differentiable terminal reward.

  20. arXiv cs.RO62

    MANGO-Grasp: A Cross-Embodiment Dexterous Grasping Framework Based on Geometry-Oriented 3D Gaussians and Mahalanobis Fields

    MANGO-Grasp proposes an anisotropic interaction framework that represents objects with geometry-oriented 3D Gaussian primitives to preserve local surface geometry, encodes hands as pretrained descriptors that capture both morphology and kinematics, and models hand-object contact through anisotropic Mahalanobis fields.

  21. arXiv cs.RO64

    FACT post-training method improves VLA success rate in contact-rich manipulation tasks

    The paper investigates why vision-language-action (VLA) models fail in contact-rich manipulation tasks that require precise physical interaction, and proposes two failure modes: precision failures (policies struggle to produce the small, precise corrective actions needed during contact) and force failures (policies fail to respond to contact forces).

  22. arXiv cs.RO55

    Diagnosing Compositional Generalization in Sequential Robot Tasks

    This paper studies compositional generalization in sequential robot manipulation tasks from the perspective of instruction-space coverage, decomposing the generalization gap into three sources: marginal instruction shift, instruction-compositional shift, and context-action shift. Results show that exhaustive instruction-tuple enumeration is unnecessary; a structured subset covering action-relevant dependencies can achieve strong out-of-distribution performance. Sparse-training failures often stem from instruction steering rather than missing low-level skills, and fine-tuning with just one demonstration per task raises out-of-distribution success from 0.4% to 54.7%.

  23. arXiv cs.RO68

    PRISM: Improving Robot Motor Control Policy Learning with Polynomial Representations

    PRISM (Polynomial Representations for Interaction Structured Motor Control) is a physics-inspired policy representation method that learns multiplicative interactions among variables by recursively adding scaled bilinear features. It only requires replacing the first affine layer of a multilayer perceptron (MLP) and can be used for reinforcement learning and imitation learning without changing training objectives, sensing, or controllers.

  24. arXiv cs.RO62

    OptCar: Specializing a Generalist FKD Model into a High-Speed MPC-Specific Model with Limited Real-Vehicle Data

    The paper proposes OptCar, a method that specializes a generalist forward kinodynamic (FKD) model into a vehicle-specific model while preserving cross-terrain generalization. At its core, a FiLM mechanism summarizes recent state-action history into a single dynamics context token that conditions multi-step predictions in a Transformer FKD architecture; specialization then combines limited real-vehicle data with targeted synthetic trajectories generated by system identification.

  25. arXiv cs.RO48

    EDAR: Learning Environment-Dependent Action Representations for Robotic Manipulation

    EDAR (Environment-Dependent Action Representation) anchors action tokens in both executable control structure and expected visual consequences. By coupling motor commands with their environment-conditioned effects, it enables the learned action space to capture interaction semantics rather than merely command-level patterns.

  26. arXiv cs.RO62

    NativeMEM: A Native Memory Compression Method for Long-Horizon Robotic Manipulation

    NativeMEM is a vision-language-action (VLA) policy with long-horizon, real-time updated memory. Its core is the Native Memory Compression encoding scheme, which reuses the VLA's own vision encoder to compress each historical frame from each camera view into a single token. Appended to the input sequence, these tokens let the pretrained VLA attend to long-term history with negligible latency overhead, requiring neither an external planner nor a newly initialized memory module.

  27. arXiv cs.RO55

    Research on Asynchronous Upper-Body Task-Space Trajectory Tracking Strategies for Humanoid Robots

    The paper proposes an asynchronous upper-body task-space trajectory tracking framework for humanoid robots, addressing the temporal asynchrony between sparse, low-frequency trajectories output by high-level planners and the high-frequency operation of whole-body controllers. The student policy is initialized through teacher-student distillation, conditioned on the fully cached future trajectory and an execution-time index, and trained with a sliding-window global reward to reduce frame drift; in the task post-training stage, an MPC module completes sparse references into floating-base and upper-body guidance.

  28. arXiv cs.RO56

    CoRL 2026 paper proposes language feedback policy that can steer frozen VLAs harmlessly

    The paper proposes a framework that interactively searches for language sequences that improve closed-loop VLA task performance, distills them into a test-time language feedback policy (LFP), and trains an improvement head to predict when language steering will be effective, then applies conformal calibration to the improvement head to prevent harmful interventions.

  29. arXiv cs.RO55

    Perceptive BFM: A Perception Control Framework for Adapting Human Motion Priors to Robot Terrain

    The paper proposes Perceptive Behavior Foundation Model (Perceptive BFM), a terrain-aware whole-body control framework for humanoid robots that uses local terrain observations from the robot's perspective to adapt contact, posture, and timing while retaining the original kinematic reference as the behavioral interface.

  30. arXiv cs.RO58

    OrbiSim: Recasting World Models as a Differentiable Physics Engine for Embodied Intelligence

    OrbiSim proposes a robot simulation paradigm that recasts world models as a fully differentiable physics engine, connecting structured scene assets, neural dynamics, and downstream reinforcement learning. The entire simulation loop, from explicit state transitions to visual observation generation, supports end-to-end differentiability, enabling tasks such as differentiable contact modeling, gradient-based policy optimization under sparse rewards, and physical reasoning. Experiments show that OrbiSim significantly outperforms existing world models in prediction fidelity and control performance, and responds consistently to asset configurations and physical parameters.

  31. arXiv cs.RO62

    OxyGen: Unified KV Cache Management for Multi-Task Parallel VLA Inference

    OxyGen proposes an inference design for unified KV cache management that treats the KV cache as a shared resource across tasks and control cycles, eliminating redundant prefill of shared observations through cross-task KV sharing and advancing unfinished language requests within each frame's decoding budget via cross-frame continuous batching.

  32. arXiv cs.RO44

    Gaussian Process-Based Robot Skill Learning and Adaptation Framework

    A study published in IEEE RA-L 2026 and presented at IROS 2026 proposes a robot skill adaptation framework based on Gaussian processes (GPs) and sparse via-points, which preserves the kinematic characteristics of skills by using first- and second-order analytical derivatives of trajectory poses. The framework includes three adaptation methods: an optimization agent, a behavior cloning agent, and a reinforcement learning (RL) agent. In simulation and real-world evaluations on three tasks—opening a drawer, pushing a cube, and manipulating a horizontal bar—it outperformed all baseline methods in success rate and demonstrated high cosine similarity and low velocity magnitude error.

  33. arXiv cs.RO50

    TRAIN: A reinforcement learning framework for terrain-aware interactive navigation for quadruped robots

    TRAIN proposes a two-stage deep reinforcement learning framework that, through the TRAIN-Loco and TRAIN-Nav pipelines, enables quadruped robots to achieve terrain-aware navigation in unstructured terrain and human-populated environments. Benchmarks compared quadruped locomotion and robot navigation baselines, and hardware experiments verified its sim-to-real transfer potential.

  34. arXiv cs.RO38

    Multi-Agent Reinforcement Learning-Based UAV Path Planning for Obstacle Avoidance in Stochastic Environments

    This study proposes a centralized training with decentralized execution method based on multi-agent reinforcement learning for UAV obstacle-avoidance path planning in stochastic environments, improved from the model predictive control (MPC) idea. Each agent communicates only with the centralized planner, plans feasible paths through extended observations based on a distance-weighted mean field approach, and uses multi-step value convergence to improve training efficiency. Comparative, ablation, and real-machine experiments verified the method's effectiveness and generalization performance.

  35. arXiv cs.RO71

    Self-Building Labs: An automated system that builds and runs laboratories according to scientific goals

    The authors propose the self-building laboratory (SBL) system, which can design, configure, and operate laboratory setups based on high-level scientific goals. Starting from an empty room or an existing self-driving laboratory, it fills in the laboratory layout, digital twin, robot skills, and experimental workflows. The system was demonstrated on a variety of robots and setups, and can scale from a single workstation to a laboratory with multiple robots working together. Its end-to-end capabilities were validated through a real-world quantum-dot synthesis experiment. The authors say this shifts the role of human researchers from manually configuring and programming device automation to specifying scientific goals.

  36. arXiv cs.RO44

    Structural Affinity: A Regularization Algorithm for Cage-Based Shape Design

    Structural Affinity is an explicit regularization method for cage-based deformation (CBD). It uses a two-cage structure and an affinity matrix precomputed with mean-value coordinates to couple the motion of inner and outer control cages, so that control handle displacements propagate smoothly to neighboring handles, greatly reducing and often eliminating cage self-intersection and mesh distortion, while adding only linear overhead during deformation.

  37. arXiv cs.RO59

    EmbodiedMidtrain: Bridging the gap between vision-language models and VLAs via mid-training

    The paper proposes EmbodiedMidtrain, which bridges the gap between vision-language models (VLMs) and vision-language-action models (VLAs) through mid-training. The authors first characterize the data distribution gap between the two, finding that VLA data occupy compact regions largely separated from the VLM distribution. They then build a mid-training data engine that uses a lightweight learnable proximity estimator to select the candidates most aligned with VLA from a large VLM pool, producing the EmbodiedMidtrain-1M dataset for mid-training VLMs.

  38. arXiv cs.RO70

    Blind assembly strategy: peg-in-hole assembly of diverse parts using proprioception alone

    This study trains blind assembly policies that observe only joint positions, velocities, and a short history of their own actions, without using visual, force, torque, or tactile sensing, and uses reinforcement learning in simulation to complete peg insertion, gear meshing, and nut tightening. The peg insertion policy transfers zero-shot to the real world and successfully inserts pegs of different shapes and sizes in 73% of trials; controlled experiments show that the blind policy generalizes better to unseen part shapes and sizes but may be slower.

  39. arXiv cs.RO58

    PRIME: Learning Precise Progress Estimation from Imprecise Demonstrations

    The paper proposes PRIME, a progress estimator for generating dense reward signals in reinforcement learning for robotic manipulation, learning precise task progress estimation from imprecise demonstration data. PRIME builds a weighted graph over observed states, connecting states across trajectories based on spatial proximity and visual consistency; edge weights encode spatial distance as transition cost, and shortest-path search approximates the geodesic distance to success, whose changes can distinguish progress, stagnation, and regression without assuming that progress is monotonic over time.

  40. arXiv cs.RO58

    Learning Robot Metareasoning: Adaptive Computation Budgets for Real-Time Control

    The arXiv paper Learning Robot Metareasoning proposes modeling the determination of the optimal computation budget as a semi-Markov decision process (SMDP), in which a single decision jointly determines the robot's action commitment duration and the time allocated to the solver. The authors trained a reinforcement learning policy conditioned on the current state, the executing plan, and low-cost planner statistics. Across three task types—ground robot navigation, agile quadrotor flight, and contact-rich manipulation—the policy adjusts its budget according to scene demands and outperforms fixed-budget baselines.