The study tested motion-tracking policies for a 24-joint humanoid in an independent CPU MuJoCo environment and found that a constant 40 ms command delay reduced the average survival rate from at least 26/30 under training randomization to 15%, with a delay margin of only 20–30 ms.
The research team deployed a VLA model on a Unitree Go2 quadruped equipped with an AgileX Piper robotic arm, achieving closed-loop mobile manipulation, and reported choices such as model variant, action horizon, and demonstration conventions that made real-robot deployment feasible. For the pick-and-place task, expanding demonstration coverage reduced held-out end-effector error by 30%, and adding secondary skills such as pushing introduced new behavior modes without degrading baseline performance. Closed-loop real-robot experiments showed that dataset composition affects task success.
RoboCode learns short-horizon motion patterns that can be reused across tasks through a shared soft-gated primitive action interface, representing action chunks as soft compositions of multiple learned primitives rather than a single discrete code. The interface is learned on action chunks pooled across multiple tasks and then frozen, after which task-specific state encoders are trained to map into the shared latent space. On five ManiSkill tasks, it achieves an average closed-loop success rate of 75.6%, above ACT's 68.8% and Diffusion Policy's 41.6%, with an average action chunk reconstruction MSE of 0.0187.
Research shows that sparse autoencoder (SAE) representations of OpenVLA activations contain outcome-predictive information and can predict task failure before the robot reaches an unrecoverable state. The authors also identify individual SAE features associated with specific stages of success or failure, and uncover recurring behavioral patterns by localizing temporal intervals where these features show significant activation changes. The team further builds a VLM-based interpretation pipeline to generate behavioral descriptions of these patterns and validates their correspondence to feature activations on held-out intervals.
RAISE-Vision is an automated platform for contact angle (CA) measurement on curved hydrogel surfaces, intended for in vitro wettability analysis of contact lenses. Its 3D-printed lens mounts, camera holders, and drying apparatus enable an automated, repeated sessile-drop workflow within a liquid handling system. A custom open-source image processing script combines elliptical fitting and the tangent method to handle droplet deformation on curved surfaces. In 10 consecutive measurements of a senofilcon A lens in air, the contact angle rose from 62.6±5.5° to 83.4±2.1°, with no significant difference from the manual protocol results (p=0.470).
The paper formulates robot-assisted dressing as a geometry-aware insertion problem and proposes a reinforcement learning framework combining a staged curriculum with a finger-wise insertion soft gate. In post-training evaluation on the deformable bracelet task, the full method achieved a success rate of 84.7±1.2%, with the insertion gate contributing the most and the highest score when combined with the curriculum. On a dummy hand, the same policy succeeded in 15 of 20 real-robot trials.
TravSUITE is a traversability system for high-speed off-road navigation, consisting of a voxel mapper that uses visual foundation models to build a geometric-semantic local map and a unified neural network that jointly predicts geometry, semantics, speed, and cost in bird's-eye view (BEV).
This paper investigates the coupling between numerical solver depth and execution horizon in flow-matching vision-language-action (VLA) models, finding that coarse-grained numerical integration reduces accuracy under rapid replanning but instead improves task performance under extended open-loop execution.
The study integrates Test-Time Training (TTT) into InternVLA-A1 and Pi0.5, training on paired demonstrations and robot trajectories to explore action generation conditioned on human demonstrations. At the current scale, adding TTT and training on paired data did not produce consistent target selection based on demonstrations, and shorter sequences, full-trajectory gradient connectivity, and contrastive regularization also did not clearly improve responsiveness to demonstrations. When demonstrations were reduced to single-frame cues, real-robot execution still preferred cups at fixed positions.
DiscoDemo uses a latent-conditioned reinforcement learning generator with privileged information, together with a parallel reverse curriculum and a success-gated diversity objective, to produce multiple successful closed-loop behaviors; after safety filtering, these are exported as a static dataset for downstream visuomotor imitation learning. On four manipulation tasks, the average downstream success rate in simulation rose from 40% for the strongest baseline to 62%, and on real robots from 39% to 78%. The authors believe that the diversity of successful solutions can improve imitation learning for pretrained flow-matching VLAs and complements corrective data collection.
Researchers propose a standardized robotic module that combines a magnetic docking interface with two actuated joints and can fold from a flat two-dimensional configuration into three-dimensional structures. Modules of a single design can be assembled into two morphologies: a three-module "trilobite" and a five-module "seal". Using a small set of discrete motion commands, an operator steers an assembled parent robot to dock magnetically in sequence with spare modules, assembling a second, structurally identical individual that then separates from the parent. The parent retains all of its original modules, forming budding-style physical reproduction.
To address input bias in GP regression caused by mobile sensor pose estimation errors, a bias-aware Gaussian process (Ba-GP) framework is proposed. It uses the recorded GP posterior-mean gradient to propagate input bias and noise into mean and covariance corrections, and provides a numerically efficient first-order correction algorithm based on the additive covariance structure.
The study proposes a modular framework that combines neural cellular automata (NCA) with existing site selection and collision-aware control for swarm robot morphogenesis. The framework introduces a decentralized attachment-site discovery mechanism in which robots already in place identify new attachment sites by updating and sharing local NCA states, eliminating the need for a predefined site list during deployment. The authors also propose offline sparse-retention fine-tuning and a false-positive loss to bridge NCA growth and gradual physical assembly. Simulations scale to several hundred robots and improve morphology accuracy, and experiments with Crazyflie quadrotors validate the integrated execution pipeline.
This paper investigates whole-body proximity sensors as an inductive bias for learning reflexive collision-avoidance behavior, finding that low-resolution lossy sensors were a stronger bias than high-resolution full coverage. The study compared eight candidate representations in a reinforcement learning environment, including latent compression, future contact anticipation, lossy corruption, and privileged state estimation, and trained over 200 humanoid robots end-to-end to play a simplified version of dodgeball under intentionally constrained sample and compute budgets. The authors also found that anticipating when and where contact occurs actually reduced learning sample efficiency.
The study used reinforcement learning to train a single policy shared by a set of twenty robots, with only the mean temperature of the five coldest robots at the end of the night as a terminal reward, and two behaviors emerged: aggregation and rotation between the core and the edge. The minimum stored energy required per robot for survival was measured at 155.6 +/- 1.3 pcm-h (1 pcm-h is 7.2 kJ of stored heat), 37% lower than stationary huddling and within 13% of a physics-based lower bound; under an energy budget in which stationary huddling would lose six robots, all robots survived.
The paper points out that when VLA models are fully fine-tuned in bf16 without keeping an fp32 master copy, AdamW rounds away any update smaller than 1/512 of a weight; 9 of 12 open-source VLA codebases train this way by default or in documented recipes, including LeRobot's SmolVLA default configuration and the π0/π0.5 guides.
Why it matters: The original quantifies the proportion of updates lost in bf16 fine-tuning and provides a LIBERO success-rate comparison for the fp32 master copy, which can serve as a reference for training configurations.
This paper proposes BM2PC, a Bayesian multimodal MPC framework for dynamic manipulation of deformable linear objects (DLOs). It formulates multimodal planning as a set optimization problem that balances control quality and behavioral diversity, and maintains multiple candidate control modes through incremental updates.
EgoForesight is a vision-language-action framework that jointly leverages robot demonstration trajectories and noisy egocentric videos for policy learning, reducing the impact of unreliable action labels through reliability-conditioned corruption and reliability-weighted regression, and supplementing supervision from demonstrated outcomes via privileged-future representation distillation.
FORCA is a coding agent that takes visual observations and wrist force/torque measurements as input, writes Python programs for contact probing and force-responsive motion, and between executions uses force records together with visual outcomes to diagnose failures and revise motion and force-feedback control logic.
WorldMirror is an agentic real-to-sim framework that builds interactive digital twins from in-the-wild robot demonstrations. It frames real-to-sim construction as the iterative refinement of an executable world program: a coding agent decomposes the scene, plans reconstruction steps, and dispatches image and 3D generation tools, then revises the program based on verification feedback. Experiments on DROID show that the framework can reconstruct visually faithful, geometrically consistent interactive worlds while preserving the physical validity of the recorded robot trajectories.
The authors propose the SEVI framework, which uses an evidential safety critic to simultaneously estimate future constraint-violation risk and epistemic uncertainty in a single forward pass, judging before an action is executed whether it is unsafe or insufficiently supported by evidence, and overriding that action with an offline-trained recovery policy.
PrefO2O proposes an offline-to-online latent preference transfer framework that enables zero-shot preference-based reinforcement learning (PbRL) on target tasks without any human feedback. The method learns a task-invariant latent preference space from a small amount of offline preference data on source tasks, disentangles transferable preference structures from task-specific variations, and uses confidence-weighted optimal transport alignment to generate pseudo-preference supervision for online reward adaptation.
KiRoSe proposes segmenting demonstration data using robot kinematic data rather than multi-camera video streams, compactly representing recurring motion events through Sparse Motion Tokens (SMTs). The method supervises an open language model with subtask annotations to interpret token sequences, clearly outperforming baseline segmentation methods on the F1 metric, and can transfer zero-shot to unseen robot embodiments and first-person human motion. Beyond segmentation, KiRoSe can also track task progress and detect action reversals on unseen robots.
The authors propose RepresentNav, a navigation foundation model that uses Representation Residuals, a three-layer transformer, to aggregate full-depth features from multiple vision foundation models and construct a representation space driven by the policy's own objectives.
Why it matters: The paper proposes constructing a visual representation space using the policy's own objectives and predicting future scene representations within it, offering transferable ideas for the representation design of navigation policies.
On October 22, AgiBot officially open-sourced the fourth installment of the AGIBOT WORLD 2026 dataset, the "Physical Reasoning" dataset, which it describes as the first real-world Physical Agent dataset.
The study proposes The Trinity, a recursive self-improvement framework that couples world generation with Harness evolution through three roles: Creator, Explorer, and Evolver.
OSWAM proposes a world action model with one-step action generation. By constructing imagination tokens to initialize the diffusion process and providing physical priors and action-conditioned representations, it compresses the tens of denoising steps required by direct-policy world action models into a single step.
PhysEvolve proposes a verifier-grounded self-evolving framework that lets vision-language models learn causal physical reasoning from executable world programs they propose themselves and from simulation-generated tasks. In each round, a shared VLM proposes world programs and evidence schemas; a deterministic physics engine freezes each valid world and simulates matched baseline and body-elided counterfactual branches to generate discrete visual tasks; then exact verifier feedback updates the policy and curriculum via GRPO.
Rollback proposes a policy improvement paradigm that turns an action-conditioned world model from a one-way simulator into a backtrackable counterfactual engine: it returns to key decision points, branches to try different actions, and compares the outcomes, thereby distilling counterfactual results into action-level supervision without learning a separate value model.
Why it matters: The world model is used as a backtrackable counterfactual engine to provide fine-grained action-level supervision for policy improvement, and the improvement margins on three real-robot tasks are reported.
The study proposes OSMA-Bench, an obstacle-conditioned safe motion adaptation benchmark built on spatially matched scenario sets that differ only in obstacle type, and introduces the Semantic Sensitivity Index (SSI) to measure whether a robot's clearance and its distance-dependent speed and smoothness follow the caution ordering of obstacles.
ETH Zurich's ANYmal quadruped robot solved a 10-pound giant Rubik's Cube using only its feet, three times, taking about 8 to 9 minutes each time. The demonstration combines computer vision, a traditional Rubik's Cube solving algorithm, and reinforcement learning to teach the robot to balance, apply force, and correct errors. This "pedipulation" (manipulation with feet) could enable future robots to open doors, turn valves, or move obstacles without hands.
DEEP Robotics has released five research results built around its self-developed embodied brain, DEEPRobotics-Brain: the world model DR-WM-1.0, the embodied foundation model DR-VLA-2.0, the embodied agent DR-Agent, the policy optimization framework DR-APO, and the data closed-loop platform DR-Benchmark.
Ding Ning's team at Tsinghua University, together with researchers from Beijing University of Posts and Telecommunications, Peking University, and Nature Will, built HandsOn, a simulation platform and evaluation benchmark for cross-embodiment dexterous manipulation, containing 66 manually verified tasks and 18,000 human simulation teleoperation demonstrations.
Why it matters: Readers can learn why dexterous manipulation evaluation should distinguish execution from transfer, and which specific failure points current multi-task policies have on bimanual collaboration tasks.
DEEP Robotics has released five research results built around its self-developed embodied brain DEEPRobotics-Brain: the world model DR-WM-1.0, the embodied foundation model DR-VLA-2.0, the embodied agent DR-Agent, the policy optimization framework DR-APO, and the data closed-loop platform DR-Benchmark. Partners include seven universities, among them MIT, Tsinghua University, and Zhejiang University.
The MonoRace system from TU Delft's MAVLab in the Netherlands, relying only on a single front-facing camera and an IMU, defeated three world-champion-level pilots in a row at the 2025 Abu Dhabi Autonomous Drone Racing competition; the paper was published in Science Robotics on October 14, 2026.
Why it matters: The Delft team broke down how a monocular camera plus IMU divides perception, state estimation, and reinforcement-learning control in high-speed racing, which can be transferred to real onboard systems with limited compute.
A research team from Harvard, NVIDIA, Berkeley, MIT and Stanford proposes RAX (Robot Agent eXploration), which connects a general coding agent to robot perception and control interfaces, uses online exploration to write successful decisions into executable programs and accumulate them as reusable skills, while keeping model parameters frozen.
Why it matters: RAX uses a frozen model plus online exploration to generate reusable programs, leading by an average of 37.4 percentage points on five simulation benchmarks with zero action demonstrations, offering a path for robot learning different from scaling data and models.
ProAS is a plug-and-play progress-aware advantage shaping framework that uses a dynamically evolving Agentic Reference Graph (ARG) to achieve fine-grained credit assignment, mitigating the problem in critic-free algorithms such as GRPO where trajectory-level advantages are broadcast equally to all actions.
The research team proposes the RynnVLA-Latent framework, which uses RynnLAM to learn continuous, 3D-grounded latent actions from visual changes without requiring native robot control signals. The method processes 30.1K hours of human, robot, and simulation video, generating 745.52M latent action descriptors, then uses flow matching to predict latent action chunks and adapt them to specific embodiments.
Why it matters: This work converts heterogeneous interaction videos into 3D latent action supervision, offering a reusable data pathway for VLA pretraining that bypasses the scarcity of action annotations.
HuGe is a family of humanoid generalist vision-language-action (VLA) models pretrained on over 2,000 hours of egocentric human and robot data. It predicts latent motion tokens to drive a pretrained whole-body controller and outputs hand-control signals.
A Microsoft team proposed the Code Video Model, in which a Coding Agent turns a prompt into a Three.js program, unifying geometric layout, object motion, camera trajectories, and event timing into an executable scene representation, then uses a program-previewed video to guide video generation.