CoRL 2026 Spotlight 🌟 A team from Alibaba, RUC, SJTU, BIGAI, and BUAA introduced Ego2Robot. It turns first-person human videos into robot training data. 18,561 hours of synthetic robot data across 15 robot morphologies. Ego2Robot tackles 3 challenges: - Turning human hand movements into robot end-effector trajectories. - Replacing human arms with robot arms in the original scenes. - Filtering out low-quality or inconsistent samples. It also supports two types of input: • Videos with existing hand-pose annotations. • Raw videos, where AI estimates hand poses and splits long videos into individual tasks. On RoboTwin 2.0’s Randomized test, the robot-only model scored 50.9%. Adding Ego2Robot data at a 1:1 ratio with real robot data raised the score to 53.5%, a 2.6 percentage-point gain.

CoRL 2026 Poster 🌟 Why can adding more first-person human video actually make a robot worse at a task? Georgia Tech researchers found a problem with traditional behavior cloning. Humans and robots see the world differently, have different bodies, and move in different ways. When human videos are added to training, models can end up learning human motions that robots simply can't perform. The team introduced EgoWAM to address this. Instead of only predicting actions, it also trains robots to predict how the world changes after an action. The team compared three ways to represent those changes: • Pixel: Predicts future frames at the pixel level. • DINO: Predicts semantic features extracted by a vision model, focusing on objects and scene content. • 3D Motion Flow: Predicts how objects move in 3D space, separating camera motion from actual changes in the environment. The team tested EgoWAM on three real-world bimanual tasks: placing a cup on a saucer, bagging groceries, and folding clothes. The experiments involved 1,800 real-robot trials. The results: Pixel prediction offered limited gains. DINO delivered the strongest OOD generalization. Compared with robot-only training, it improved success rates by: • 43 percentage points on cup placement. • 40 points on grocery bagging. • 70 points on clothes folding. Clothes-folding success in OOD tests reached 85%. 3D Flow was better for precise spatial manipulation. It improved in-domain success rates by: • 20 percentage points on cup placement. • 30 points on grocery bagging. Both tasks reached 90% success. On grocery bagging, 3D Flow also delivered a 45-point gain in OOD testing. EgoWAM points to a different way of training robots. Authors: Baoyu Li, Xinchen Yin, Mengying Lin, Yixin Zhang, and Danfei Xu.

CoRL 2026 Oral 🌟 A team led by Peking University introduced OpenDexGrasp. It takes free-form language instructions, multi-view images, and 3D point clouds as input. Then it directly generates a dexterous five-finger grasp pose. For example, grip a hammer by its handle, not its head. To train the model, the team built OpenDexVerse: • OpenDex-Scale: 1.24M automatically generated grasp samples. • OpenDex-Align: 27.42K human-teleoperated and transferred grasp samples. The results: In simulation: 62.96% functional grasp success on unseen object categories. The DexGraspNet 2.0 baseline scored 43.22%. On a real robot, it achieved a 72% success rate across 10 everyday objects, including brushes, shampoo bottles, umbrellas, and pitchers. The baseline scored 59%. It's also faster. Traditional methods predict the graspable area first, then optimize the hand pose. OpenDexGrasp generates the grasp pose directly. Traditional pipeline: 2.84 s OpenDexGrasp: 0.93 s 67% less inference time. It's not perfect, though. Thin or complex objects like keyboards, frisbees, and flower pots can still cause problems, including finger penetration and inaccurate grasp placement.

CoRL 2026 Poster 🌟 A team from UC San Diego developed HuG, a framework for humanoid robot grasping. Their goal: Enable humanoid robots to find objects on their own and grasp them stably through coordinated whole-body movements, without relying on real-world grasping demonstrations. They break the task into three stages: Perceive: Find a viewpoint that reveals the target. Align: Position the whole body for a feasible grasp. Grasp: Reach out and pick up the object. The key is how HuG learns which grasps are actually executable. The team put the same whole-body controller used on the real robot into the simulation loop, filtering out grasps it couldn't reliably execute. 73% of grasps that were kinematically valid failed this test. HuG ran on a Unitree G1, trained entirely in simulation with zero real-world demonstrations. The results: • 87.5% success on a real-world multi-level shelf task. • 85% on objects unseen during training. • 73%+ across all five tested table heights in a separate experiment.

From Tesla Optimus to Anthropic. Bingjie Tang has joined Anthropic as a Member of Technical Staff. At Tesla, she worked on dexterous manipulation for the Optimus humanoid robot, covering data collection, policy training, evaluation, and real-world deployment. Before that, three years at NVIDIA Seattle Robotics Lab with Yashraj Narang and Dieter Fox. Her research focused on contact-rich manipulation and sim-to-real transfer. A few of her projects: • IndustReal (RSS '23): Transferring contact-rich assembly skills from simulation to the real world. • AutoMate (RSS '24): A dataset of 100 assemblies, with specialist and generalist policies that transfer from simulation to real robots. • SRSA (ICLR '25 Spotlight): Retrieving and adapting existing robotic skills for new assembly tasks. Her academic path: HUST (BS) → Brown (MS) → USC (PhD), advised by Gaurav Sukhatme.

Ant Group just open-sourced LingBot-Touch, a tactile foundation model for robots. Cross-sensor transfer, even to sensors unseen during pretraining. Nearly 5k hours of real-world tactile data. Tested on 26 tasks, 2 robotic hands, and 4 action policies. Average success rate: +14 to 26 percentage points over no touch. Bottle-cap twisting: 56.25% with no touch. 62.5% with raw tactile input. 90% with LingBot-Touch features. Two sizes: Tiny & Small Apache 2.0. Weights now available.

⚡️Alibaba just launched QwenWork 2.0 Teams can now work alongside AI agents in a shared workspace. • Team Collaboration: Bring a digital employee into team workflows. It can connect multiple DingTalk chats, docs, AI spreadsheets, and meeting notes, while handling parallel sessions and assigned tasks. • Automated Tasks: Set event-based triggers for messages from specific people, @ mentions in selected groups, and approval requests. The agent acts automatically when conditions are met. • Human-AI Co-Creation: Review content and leave feedback directly in chat. AI can apply multiple edits to the original document, spreadsheet, slide, or app. • Resource Space: Collect AI-generated and uploaded files in one place, with support for browsing, searching, and sharing.

ByteDance tested Seedream 5.5 with 100 consecutive edits of the same person. Each round used the previous output as the starting point. As clothing, accessories, and tabletop objects changed, Seedream 5.5 could still make precise edits to tiny details like rings and eyeglass frames. It also handled how accessories interact with the skin, including contact, occlusion, and lighting. Here's the demo 👇

ByteDance just launched Seedream 5.5🚀 Outperforming GPT Image 2.5 Sunburst & Nano Banana 2.1 in both text rendering fidelity and visual realism. What's new: • Supports continued work on the same image across 100 rounds of editing. • Native transparent/translucent background export (Alpha channel) • Doubled usability for cinematic and office use cases vs. 5.0 Pro Seedream 5.5 is gradually rolling out across Dreamina, Doubao, Doubao Work, and Volcano Engine Ark.

A Yale CS PhD candidate has open-sourced the Python scripts behind figures in his top-conference papers. 10.3K stars, 715 forks. It's called Figures for Papers, created by Chen Liu, whose work has appeared in Nature Machine Intelligence, ICML, NeurIPS, and ECCV. That means you can learn from real publication-quality examples, including their color palettes, typography, legends, and layouts. It covers everything from bar charts and composition breakdowns to 3D spheres, radar plots, line charts, concept diagrams, and trend plots. License: CC BY-NC 4.0, not MIT.

A team from Zhejiang University and vivo has introduced FlashRelight, an AI system that can relight portrait videos. It's more than just applying a filter. The goal is to change the lighting while preserving a person's identity, expression, and motion. FlashRelight takes a two-stage approach: First, a delighting model suppresses the original lighting and shadows. Then, a relighting model generates new dynamic lighting across the video. You can control the lighting in three ways: • Envmap sequences: Precise frame-level lighting control. • Text prompts: Describe color, direction, movement, and flicker. • Reference videos: Transfer lighting effects from another video. The team built a hybrid dataset with 34k+ paired videos: • ~24K synthesized from static OLAT lighting data • 10k+ generated by Wan2.2 This gives the model varied lighting and motion without costly dynamic light-stage capture. On envmap-based relighting, FlashRelight outperformed all baselines across all five evaluation metric For short videos, ads, or virtual-human content, this could make it easier to adjust a portrait’s lighting in post-production.

Alibaba Cloud just increased monthly quotas for all four Token Plan personal edition, with the biggest gains reaching 78% Effective Nov 9: Lite: 11.5K → 20K Credits/month. Still ¥39/month. Essential: 25.5K → 45K Credits/month. Still ¥79/month. Standard: 45K → 80K Credits/month. Still ¥139/month. Pro: 180K → 200K Credits/month, while the monthly price drops from ¥499 to ¥312. The effective cost per Credit is down by roughly 42–44% across all four plans.

DeepSeek Harness just launched Creator Mode, letting users build plugins with natural-language prompts. But developers have been going a step further. Meet dsh-purge, a project that gives developers more control over Harness prompts, permissions, sandbox policies, and tool execution. It has two main modules: Clean is the main control panel. Developers can edit system prompts, switch model-specific rule sets, manage files like AGENTS.md and CLAUDE.md, and adjust approval rules for terminal commands, file operations, and sub-agents. Unlike prompt-only tweaks, dsh-purge patches local @deepseek-ai packages and config files to change how Harness works. It also supports applying, restoring, and uninstalling patches. Drill is built for authorized local security testing, with tools for managing assets, test environments, skills, and knowledge bases. The project also backs up original files. After a Harness update, patches that no longer match the code are skipped or flagged for review rather than forcibly applied. Repository status: The original author's repository is no longer available. GitHub copies now include a team-maintained version and independent mirrors.

Turing Award winner Andrew Yao and Moonshot AI founder Zhilin Yang have co-authored a new paper. They're asking a question: Can today's frontier LLMs do real theoretical computer science research? To find out, the team built TCSExam, a new benchmark for AI research. It covers four areas: • Theorem Proving • Solver Engineering • Mathematical Constructions • Open Conjectures Unlike typical benchmarks, TCSExam doesn't just test whether a model can solve a problem. It checks whether its work actually holds up. The team put together 100 tasks for the Full test and 20 for Lite, evaluating 8 agents on Full and 9 on Lite. Every submission is checked by the Lean kernel or an executable checker. Here's what they found. On theorem proving, GPT-6 Astra hit a 70% success rate with no hints. Kimi K3.1 Pro came second at 38.3%. No other model scored above 11.7%. Constructions results: GPT-6 Astra scored 104.9%, beating the existing reference results. Kimi K3 also pushed two records forward: • Graph theory: 630 edges in a 124-vertex graph with no 3- or 4-cycles, up from 629. • Quantum error correction: Improved a code's parameters from [[25,7,5]] to [[25,7,6]]. What about problems that are still unsolved? The team tested 30 open conjectures. No model independently solved any of them from start to finish. AI isn't ready to do frontier research on its own. But it's already producing results that researchers can check and build on.

HKUDS's DeepTutor just passed 55k stars. For a study workspace. It tackles a familiar problem: a pile of textbooks and papers you can't get through, and a chatbot you don't trust to stay on the material. DeepTutor brings the whole learning process into one workspace. It combines Chat, Books, Reading, Practice, Task Board, Calendar, and a learning notebook. Create a workspace for each subject, upload your materials, and keep asking questions and practicing based on the same knowledge base. The biggest difference from regular AI chats: Your learning stays connected, with shared context and progress, instead of being scattered across disconnected conversations. It's open-source under Apache-2.0, with v1.7.1 released on November 1.

AGILINK's OmniGlove is on sale: put it on, capture your hand motion and tactile data, then drive a dexterous hand directly. It tracks at ≤1mm position and ≤1.5° pose, 10ms latency, and holds steady near metal and magnets, where most gloves drift. - Standard: motion capture + teleop - Pro: adds full-palm tactile, 746 sensing points per hand, 0.1N force resolution, <3mm spatial resolution Now on sale from RMB 29,000 a pair.

Meet China's new robot news anchors! They're jointly developed by Xinhua and UBTECH. Both robots are built to 1:1 human scale. And their looks were designed by Zhang Yimou, the filmmaker behind the opening and closing ceremonies of the 2008 Beijing Olympics.

Xiaomi's new world model can generate a seven-camera driving sequence lasting ~ 45s. It's called Xiaomi Auto World Model 2.0, a unified streaming causal world model for Physical AI. Generating long driving videos comes with a problem: scenes can start to drift over time. Xiaomi tackles this with WorldRec 2.0. It first reconstructs static and dynamic 3D Gaussians from footage captured by multiple cameras across different time steps. Then uses this 3D scene to guide future video generation, helping keep road geometry and visual details consistent over time. The results go beyond video generation. E2E-WM-RL vs. ST-JEPA-E2E in 1,000 closed-loop simulation scenarios: • Collision rate: 1.2% → 0.4%. • Scenario completion rate: 95.2% → 96.3%.