DexPort is a hand-held data collection interface for high-DoF robot hands that uses the same hand hardware for both collection and deployment, recording joint target commands and measured joint states. The operator moves the hand-held interface to specify wrist motion and coordinates finger motion by selecting or adjusting skills, without needing to control each joint individually.
ViT-H proposes a generation-plus-tracking video-to-humanoid control framework. It uses a diffusion model to generate humanoid reference motions directly from monocular video, skipping human motion reconstruction and retargeting, and uses motion-tracking results to post-train the generator with direct preference optimization. The authors also present an autoregressive variant that enables real-time humanoid video teleoperation with sub-400 ms latency. Experiments on the Unitree G1 show that ViT-H outperforms existing reconstruction-retargeting-tracking pipelines in motion quality and executability, and that performance improves with model size and with pre-training and post-training data scale.
RynnUnify is a 16B unified embodied model that combines embodied understanding with long-horizon visual planning through interleaved video-language reasoning, covering spatial reasoning, physical prediction, and grounding. The authors also developed an agentic extension based on reinforcement learning to selectively use video prediction to reason about unobserved states and interaction outcomes, and a world-action extension that translates understanding and prediction capabilities into robot control.
The study proposes an OCT-guided data-driven framework that learns soft tissue shape control from user demonstrations and executes it with closed-loop OCT feedback. In 12 configurations and 576 phantom trials, the Explicit policy using residual-conditioned voxel point clouds achieved Shape RMSE of 0.534 mm on the depth task and 0.443 mm on the directional task; full-volume Diffusion with residual conditioning reduced directional error from 0.748 mm to 0.555 mm.
The authors propose Backtrack, a lightweight failure detection and recovery mechanism for frozen VLAs that requires no recovery demonstrations, policy retraining, or additional inference models. It uses a failure prediction head that adds only 0.48-0.88% to the parameter count to read the VLA's internal representations; when the predicted failure probability exceeds a threshold, it pauses the policy, backtracks the robot arm's recent motion, and hands control back to the original policy for a retry.
HoloBoxing proposes a humanoid robot boxing learning approach built on a pretrained whole-body controller, separating reusable motor skills from task decision-making: a compact planner outputs future body-point targets, while the frozen controller handles joint-level execution.
Regarding the prospects of combining federated learning with VLA policies, the research points out that sharing human-correction adaptations across tasks, environments, people, and robot embodiments carries safety risks: a shared update that raises average task success may suppress rare safety interventions or produce unsafe responses on another robot.
The study proposes an action-aware quantification method that uses a vision-language model to decompose long-horizon manipulation trajectories into object-level subtasks and assign weighted action credits, and builds a training video dataset of more than 800 hours plus the LOOM benchmark, which contains 19,618 human-calibrated evaluation frames.
AWARE is a world-model framework that relies solely on external camera monitoring, performing anomaly detection through system dynamics estimation and state prediction without onboard sensors such as motor encoders. On 240 real-world robotic crane trajectories (including payload collisions and motor degradation), its anomaly classification improves by 26% over the strongest baseline; the inferred dynamics parameters can also be decoded to localize the source of faults, achieving 88% accuracy in localizing degrading motors.
The study evaluates whether Conditional Neural Processes can autonomously learn action representations that are informative about underlying goals. It finds that adding an attention mechanism to the latent space enables robust goal representations to form early in the observation of an action sequence. The authors accordingly suggest the model could serve as a general self-supervised learner for robots. The work was presented as a poster at the IROS 2023 Workshop on World Models and Predictive Coding in Cognitive Robotics.
The paper proposes a natural-language policy interface that connects operator instructions to ROS 2 Nav2: a local large language model converts instructions into JSON policies specifying destinations, spatial constraints, and temporal or event conditions; deterministic checks then decide whether to accept them, and three costmap adapters inject them into the existing navigation stack without changing the planner or controller.
The authors propose OWLS (Open-world Language-conditioned Speed Fields), which plans directly over the dense feature maps of a vision-language model (VLM): given a description of traversable terrain, the VLM scores every pixel of a single camera image, and OWLS projects these scores onto the ground, adapts them to the platform with two morphological operations, and uses them as the speed field of a minimum-time planner, so that terrain appearance and obstacle spacing jointly shape a single trajectory.
UAV-SwarmAgent is a hierarchical multi-agent system in which an operator issues a mission in one natural-language sentence, and a ground-station agent generates and executes code that turns it into routes for each UAV. The system uses a vision-language model to locate inspection targets on an aerial map, balances them among N UAVs by route length, and then worker agents equipped with lightweight decision models verify incidents.
The authors propose HGWM, which combines an LLM-generated room-anchor prior with persistent spatial-semantic memory to select search directions in zero-shot object navigation. On HM3D v0.1, it reports 62.6% success rate and 31.5% SPL, versus 58.1% and 31.2% for WMNav; on MP3D, 46.8% and 17.9%, versus 45.4% and 17.2%.
The v1 lecture notes "Optimization, Optimal Control, and Reinforcement Learning for Robotics," compiled and edited by Yonghyeon Lee, have been released on arXiv, emphasizing the mathematical connections and practical applications among the three. The notes were written through human-AI collaboration and are a compilation of existing material rather than new research results; they are currently a preliminary version with no figures or graphs, and revisions and expansion are planned.
This research proposes a unified control framework that shifts multi-robot systems from "obligate swarms" constrained to pure self-organization toward "facultative swarms"—hybrid architectures that operate with centralized planning when conditions are permissive and degrade smoothly into decentralized swarming under environmental stress. The authors argue for redefining swarming behavior as a dynamic capability rather than a fixed system category, thereby exploiting centralized optimality when conditions allow and the flexibility of self-organization when they do not.
The study proposes an anatomy-anchored imitation learning framework that combines anatomy-guided probe placement with learned ultrasound navigation. Across 16 unseen participants, task success rates for standard-view acquisition of the kidney, gallbladder, and liver and for radial artery scanning were 96%, 94%, 88%, and 94%, respectively.
CoCarry introduces a dataset and a contact-support-structured retargeting method for dual-humanoid cooperative carrying. The dataset contains 220 synchronized dual-person sequences spanning 21 task schemes, 5 shared-object geometries, and 4 terrain types, totaling about 61 minutes of motion data.
The authors propose R²E, a framework for automated reward function design that uses explicit role specialization to mitigate insufficient exploration and optimization instability in LLM-based reward design. The framework includes a novelty explorer responsible for expanding the search space, a guardian responsible for conservative refinements to improve training stability, and an evolver that recombines high-performing rewards through crossover and mutation, with a dynamic scheduling strategy that gradually shifts the search from broad exploration to focused exploitation. R²E also introduces a global elite pool that retains the best-performing rewards to guide subsequent iterations. Experiments on multiple robotic tasks show the effectiveness of the method.
LogLook proposes an execution-log-style memory that lets frontier vision-language model planners record executed actions and earlier notes, and selectively review image frames based on expected changes. With the benchmark's built-in motion planner, it matches the success rate of note-only memory while using 43% fewer tokens and less than half the images, and it causes notably fewer failures due to memory or observation errors; under direct control, it solves 93.5% of the RoboMME test set.
DynaSlots is an action-label-free 3D visual pretraining framework that decomposes each point cloud frame into a small set of temporally consistent slots and models dynamics at the slot level rather than the scene level. A temporal slot propagation module keeps slot assignments consistent across frames; an inverse dynamics model extracts a latent action for each slot from feature differences; and a diffusion-based forward dynamics model, paired with a self-supervised interaction head, predicts each slot's future state and learns pairwise proximity and motion relationships.
MAST-VLA proposes a dual-conditioning framework that introduces skill-conditioned LoRA Mixture-of-Experts adapters into the LLM backbone and converts predicted affordance heatmaps into a patch-level trajectory-affordance spatial prior for the action head. In the CALVIN ABC→D long-horizon multi-skill cross-scene setting, it achieves an average completion length of 4.65 over five-subtask sequences, the best performance in that setting. Real-robot multi-skill sequence experiments further validate its deployability under real perception and control noise. The paper has been accepted to CoRL 2026.
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.
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.
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.