For fine-tuning VLA models, the article presents a reproducible parameter-efficient training approach spanning from full SFT to LoRA. The content covers these two training paths for developers to reproduce according to the approach.
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.
Official weights of InsertAnything's force-conditioned insertion policy were evaluated in Isaac Lab: at 0.1 mm clearance, baseline success rates for two hole types were 96–100%. With a 5 mm initial hand offset, success rates were unaffected, but with a 3 mm hole-position observation deviation, the hexagonal hole success rate dropped to 60%.
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.
OpenCR-MuJoCo presents a discretized simulation method for tendon-driven continuum robots in MuJoCo, modeling the continuous backbone as a rigid chain with elastic joints, where bending is produced by tendon length differences. The work uses SoRoSim reference data and the official demo to validate static accuracy, release oscillation, and closed-loop tracking.
This study takes a bicycle robot with a reaction pendulum as its subject, derives the roll dynamics and the speed-dependent instability mechanism, and uses proximal policy optimization (PPO) to train coordinated control of the steering, rear wheel, and reaction pendulum. Experiments are carried out in PyBullet: natural fall, training, balancing, navigation, and disturbance recovery, concluding with obstacle-avoidance training.
The article documents unsupported operator issues encountered when running VLA models on RK3588 and compiles them into a list and troubleshooting log. The material provides only the title and summary; specific operators, model versions, or performance data are not given.
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.
TorqueAGI is providing a "world reasoning model" intelligence layer for industrial robots and heavy machinery rather than developing its own hardware. The model represents objects, machines, and workflows as spatiotemporal graphs and incorporates physics and geometry into its reasoning. The company says new tasks typically require only dozens of examples rather than tens of thousands, and must be validated in simulation and in supervised operation at customer sites before unattended operation is possible. The same model has already run on single-arm and dual-arm robots, mobile platforms, scissor lifts, humanoid robots, and tractors, while details specific to each machine body still need to be described per machine.
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.
At the 2026 AgiBot Innovation Day on October 29 at the Shanghai World Expo Center, Peng Zhihui released the third-generation hardware architecture AHA3, the foundation model GO-3 Preview, the multimodal interaction model WITA-OMNI and Lingqu OS 1.0, and proposed Worken, a new unit for measuring robot value (number of active robots multiplied by autonomous working hours multiplied by the output efficiency ratio).
Foreign media report that Physical Intelligence is in talks for a new funding round, planning to raise at least $1 billion at a target valuation of about $20 billion, more than 8 times its $2.4 billion valuation in November 2024. The company was founded in 2024, develops the π series of robot models, and opened a beta version of the Pi API on October 13. The article also compares Skild AI: 10 months of formal commercial deployment, more than 60 paying customers, ARR exceeding $100 million, and cumulative recognized revenue of $50 million.
At its AI DAY on October 29, Critical Point announced that cumulative production of its dexterous hands has surpassed 15,000 units, and launched its self-developed data-collection glove, OmniGlove Pro. The glove uses a fusion of EMF magnetic field positioning and IMU, achieves fingertip positioning accuracy of 0.1 mm, is equipped with 746 tactile sensing points, and has a tactile spatial resolution of 3 mm. The company also introduced the "real-machine Ready" concept, spanning from data collection to real-machine replay.
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.
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.
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.
The research treats complete closed-loop robot policy implementations as reusable execution experience: a coding agent generates policy code from a few successful demonstrations and iteratively improves it using simulation feedback. Across four source tasks in RoboCasa, iterative optimization raised the success rate from 28.3% with initially generated code to 64.2% with validation-selected code, and the learning approach was validated on a physical Unitree G1.
UniCross proposes a unified framework that models four dexterous manipulation skills—grasping, relocation, in-hand rotation, and in-hand translation—as the same relational-target formulation, and distills a single policy that does not depend on skill identifiers. The policy performs well on each individual skill, generalizes to unseen objects, and supports long-horizon skill chaining and execution of relational targets unseen during training; the framework also applies to different hand morphologies.
The authors propose DR-LfD, a framework that integrates visuomotor policies into a TAMP-gated decision-making system, decomposes human demonstrations into atomic skills based on contact relationships, and then reproduces and reorganizes them using visuomotor policies or object-centric primitives. This method shifts the demonstration data requirement from growing exponentially with the skill sequence to growing with the number of skill types, and on two DexMimicGen tasks it outperforms a diffusion imitation policy trained on 1,000 demonstrations while using only 100 demonstrations. The paper has been accepted by T-RO.
The authors propose SPARK (Sequential Planning via Anchored Robotic Keypoints), which describes tasks as a symbolic behavior tree, anchoring to the current 3D observation only when each action is invoked; it verifies the plan at runtime and re-anchors objects on local failure, requesting a new plan only if this step fails.
VEGA proposes a method for training navigation Vision-Language-Action (VLA) models from unlabeled egocentric navigation videos. It reconstructs local scene geometry from monocular video, samples text, image, and spatial waypoint goals, and generates obstacle-avoidance trajectories, then uses these trajectories to train a flow-matching VLA navigation policy; geometry is used only during training.
This study proposes Gaussian Evolutionary Specialists (GES), a reinforcement learning-based co-design framework. By maintaining a population of specialist policies whose design-space territories evolve according to their control performance, it mitigates behavioral diversity collapse, in which a single shared policy converges to the same control strategy across different morphologies.
NDPP-Grasp proposes injecting physical plausibility constraints directly into the denoising process of a task-aligned grasp diffusion model, so that guidance can be applied throughout generation even when the constraints are non-differentiable. Existing diffusion methods first train a grasp diffusion model for task alignment and then apply physical plausibility correction after generation, leaving the generation trajectory itself unguided by physical constraints. The work was accepted to NeurIPS 2026.
The paper proposes a specification-aware action distribution optimization framework that enforces a class of Signal Temporal Logic (STL) constraints at execution time without modifying the parameters of a pretrained robotics foundation model. At each decision step, the method reasons over the remaining horizon via forward dynamics propagation to compute a minimally modified action distribution that satisfies a hard STL feasibility constraint, and it is validated in simulation with state-of-the-art robotics foundation models across multiple environments and complex specifications.
FlowTouch is a view-invariant visuo-tactile prediction model that uses an object's local 3D mesh to encode information for predicting tactile patterns, enabling tactile feedback even before the object is touched. The method combines scene reconstruction with a Flow Matching-based image generation model. Results show that it narrows the sim-to-real gap and generalizes to new sensor instances, and the generated tactile images can also be used for downstream grasp stability prediction. Code, dataset, and videos are publicly available.
The authors propose π-StepNFT, an online reinforcement learning framework for streaming VLAs that treats transitions in SDE sampling as supervision units and uses variance-normalized transition errors to distinguish positive and negative velocity branches, without computing the final-action likelihood.
The authors propose CHOP, which fine-tunes visuomotor navigation policies with counterfactual human preference labels to align them with human intuition about safety and obstacle avoidance. On the SCAND dataset, multiple advanced models fine-tuned with CHOP reduced near-collision events by an average of 49.7%, decreased deviation from human-preferred trajectories by 45.0%, and increased average obstacle clearance by 19.8%.
WildOS is a unified system for long-range, open-vocabulary object search that combines safe geometric exploration with semantic visual reasoning for complex outdoor environments without prior maps and with limited depth sensing. The system builds a sparse navigation graph to maintain spatial memory and uses a foundation-model-based vision module, ExploRFM, to score the graph's frontier nodes, simultaneously predicting traversability, visual frontiers, and object similarity to enable real-time onboard semantic navigation.
EquiContact proposes a hierarchical framework consisting of a high-level vision planner, Diff-EDF, and a low-level compliant visuomotor policy, G-CompACT, for learning contact-rich manipulation policies that generalize across task configuration spaces.
The authors propose CWAM (Causal WAM), which adds sparse routing paths to the action head of a world action model to separate the latent variables that truly drive actions from the nuisance variables that scene prediction depends on, while keeping the world modeling objective unchanged. The paper proves that under a sparsity constraint, the latent variables used to generate actions are an invertible function of the action-relevant latent variables alone, and the true structure is identifiable up to a permutation of the latent variables, without requiring sufficient temporal dependence or assumptions about how unobserved states evolve. Synthetic data and real-robot experiments validate the theory and show improved closed-loop success rate efficiency.
FOCUS combines local optimal transport alignment with short-horizon differentiable actor-critic learning for observation-only imitation learning, requiring neither temporally aligned expert trajectories nor expert-aligned initialization. The method matches agent rollout windows to multiple sampled expert segments using a soft best-of-K objective, with the transport cost providing a differentiable imitation reward and a learned critic bootstrapping value beyond the rollout horizon.
SACQ is an ultra-low-bit weight quantization framework for Vision-Language-Action (VLA) models, combining structure-aware codebook fitting, action-aware mixed-bit allocation, and two-stage post-quantization refinement.
MotorSense is the first large-scale paired egocentric video and surface electromyography (EMG) dataset, containing synchronized video, bimanual EMG, IMU, and audio across 100 hours of unscripted manipulation. The research uses this data for cross-modal pre-training between video and EMG to learn an action organization shared across visual and motor spaces, and transfers it to tasks including action recognition, 3D hand reconstruction, and 3D contact reconstruction. Across four action-related visual tasks and eight video datasets, this sensorimotor-aligned representation outperforms conventional video self-supervised methods.