To start off this month’s digest, I’d like to highlight another new volunteer contributor that has been assisting with maintaining the COMPARE Ecosystem repositories on robot-manipulation.org: 
Oluwatosin (“Tosin”) Kolade is a graduate student in Nigeria researching space robotics. He is an active member of IEEE RAS, volunteering his time to contribute to the Student Activities Committee, as a Science Communications Ambassador for IROS 2026, and now as a contributor to the COMPARE Ecosystem, primarily reviewing open-source assets for Simulation and Learning. And as you’ll find below, Tosin has been quite busy reviewing new entries for both of these repositories. Are you interested in helping out? Let us know! 
Much of the robotics community will descend upon Pittsburgh, Pennsylvania, USA over the weekend to attend the IROS 2026 conference. If you want to learn more about COMPARE, we’ll have a presence through the week including our own workshop running on Sunday, September 27: Reproducing Robot Manipulation: Developing Guidelines to Improve Accessibility and Reproducibility of Open-Source. Members of the project team will also be assisting with running competitions, presenting papers and posters, and speaking at workshops, including one on Thursday, October 1, Reproducible Benchmarking of Robotic Grasping and Manipulation: From Advanced AI to Generalized Humanoid Intelligence, from co-PI Kostas Bekris titled, “How Will We Know Robot Grasping Is Solved?” Don’t miss it! We’re planning some website reorganization over the next month or so; keep an eye out for those announcements here. Until then, mark that bench, keep the source open, and see you next month! |
💬 #software: Intrinsic launches open-source set of ROS-compatible capabilities to accelerate the development and deployment of physical AI: Intrinsic Core [blog post] 🤖 Grasp Planners: New entry added (61 total): Dex4D is a framework that leverages simulation for learning task-agnostic dexterous skills that can be flexibly recomposed to perform diverse real-world manipulation tasks, learning a domain-agnostic 3D point track conditioned policy capable of manipulating any object to any desired pose. This ‘Anypose-to-Anypose’ policy is trained in simulation across thousands of objects with diverse pose configurations, covering a broad space of robot-object interactions that can be composed at test time. During execution, Dex4D uses online point tracking for closed-loop perception and control. 🤖 Learning Environments: New entries added (37 total): Gymnasium-Robotics: A collection of reinforcement learning robotics environments built on the Gymnasium API, running on the MuJoCo physics engine via its maintained Python bindings. Bundles several environment families (Fetch, Shadow Dexterous Hand, Maze, Adroit Hand, Franka Kitchen, MaMuJoCo) under a single package, plus a multi-goal GoalEnv extension of the core API for goal-conditioned RL (e.g. Hindsight Experience Replay). Isaac Lab-Arena: An open-source extension to NVIDIA Isaac Lab for composable environment creation and generalist robot-policy evaluation at scale. Environments are assembled on-the-fly from three independent, reusable primitives — Scene (object / furniture layout), Embodiment (robot + observations / actions / sensors / controllers), and Task (objective) — via an ArenaEnvBuilder, eliminating the combinatorial-config duplication of traditional task libraries. RoboLab: A robot- and policy-agnostic simulation benchmark, built on NVIDIA Isaac Lab, for evaluating real-world task-generalist manipulation policies (policies trained on real-world data, not co-trained in sim). Scenes are built by physically arranging objects in simulation, tasks are created by attaching language instructions to a scene, and both can be generated by hand or via AI agentic workflows. Robust Gymnasium: A unified, modular benchmark for robust reinforcement learning, built around a "disruptor module" formalism (an MDP augmented with perturbations to observed state, observed reward, action, or environment/dynamics). Covers 170+ tasks spanning control, robot manipulation, dexterous hands, safety, and multi-agent settings, with perturbations that can be random, adversarial, arbitrarily-set, or semantic/domain-shifted, applied at any point during training or testing. Also supports using LLMs as adversarial policies to set robust parameters. SoftMimicGen: An automated data generation pipeline (in the MimicGen family) for deformable object manipulation. From a small set of human teleoperated demonstrations, it registers and transfers subtask segments onto new object/scene configurations to synthesize large datasets — extending the "MimicGen"-style rigid-body data generation paradigm to deformables (cloth, rope, tissue, stuffed toys). Generated data is used to train policies that show zero-shot sim-to-real transfer, with further gains from sim-real co-training. Tektonian Simulac: Simulac is a cloud-based robot benchmark execution platform that enables users to run and evaluate robot learning models without local simulator setup. It takes a VLA model along with benchmark configurations as input, and executes benchmark episodes in parallel to produce evaluation results such as success rates, trajectories, and logs. By removing environment setup and enabling fast, reproducible runs, Simulac simplifies and accelerates the process of benchmarking robot learning systems. 🤖 Simulators: New entries added (21 total): GSWorld: A photo-realistic simulator for robotic manipulation that fuses 3D Gaussian Splatting (3DGS) with physics engines to close the sim-to-real visual gap. Introduces "GSDF" (Gaussian Scene Description File), an asset format that combines Gaussian-on-Mesh scene representations with robot URDFs, letting the same environment be used to train, evaluate, diagnose failures on, and relabel manipulation policies in a closed loop. OmniSim: Open-source robotics simulator built for coding agents to control via HTTP/JSON and MCP; uses Newton as its sole physics backend, wgpu rendering, and a ROS 2 sidecar (simulation_interfaces standard). Positions itself as agent-authored: harness, physics integration, cloth/soft-body stack, RL pipeline and ROS 2 sidecar were written by an AI agent under human direction. |
🤖 Tools and Testbeds: New entries added (17 total): EgoInfinity is a modular, web-scale 4D hand-object interaction data engine that converts in-the-wild monocular RGB video (with text descriptions) into metric hand trajectories, object geometry/pose, and contact-relevant states — with no manual annotation, wearables, mocap, or CAD required — and retargets the recovered motion onto diverse robot embodiments for video-to-action learning. Rather than a static dataset, it's an upgradeable pipeline (perception → segmentation → reconstruction → interaction-aware refinement → retargeting). RoboChallenge is the first large scale real-robot-based benchmarking of embodied intelligence. Robots are moving into the real world, but there is still no unified, open, and reproducible benchmark. The advantages of RoboChallenge include diverse tasks (from object manipulation and scene interaction to long-horizon planning), multiple robots (supporting various robot morphologies), open and fair (all results and rankings are displayed transparently on the platform). RoboEval is a structured benchmark for bimanual manipulation, featuring diverse tasks with varying coordination and complexity. Unlike existing benchmarks that evaluate policies solely based on task success, RoboEval introduces an initial suite of tiered, semantically diverse manipulation tasks with fine-grained diagnostic metrics to probe the capabilities and failure modes of learning-based agents. The benchmark provides 8 task families with 28 total variations that target specific skills such as coordination, precision, and interaction under variability, and are accompanied with 3,000+ total human-collected demonstrations. VLA-Replica is a low-cost, easily reproducible real-world benchmark for evaluating VLA models. Built from off-the-shelf components, the system can be quickly assembled and replicated across laboratories, providing a consistent environment for policy evaluation anywhere in the world. VLA-REPLICA includes a diverse suite of manipulation tasks and a small-scale demonstration dataset for target-domain adaptation, with real-world evaluation protocols for both in-distribution and out-of-distribution settings. |
💬 #events: NVIDIA launches the Video2Data Challenge, focusing on advancing Real2Sim capabilities for robotics [website] What’s coming up soon?IROS 2026: IEEE/RSJ International Conference on Intelligent Robots & Systems (IROS) 2026, September 27 - October 1, 2026, Pittsburgh, PA, USA 
CoRL 2026: Conference on Robot Learning (CoRL) 2026, November 9 - 12, 2026, Austin, Texas, USA 
|
