Realtime AI News
Guanglun Zhineng releases RoboFinals RL Envs for long-horizon robot training
Guanglun Zhineng has released RoboFinals RL Environments, an interactive reinforcement-learning simulation environment for robot models, unveiled during CoRL 2026. Its first demo is an automotive repair task lasting more than eight minutes, built on the company's SimFoundry simulator and paired with a matching benchmark.
As large models move into tool calling, software operation and multi-step tasks, the centre of gravity in training is shifting from data and answers toward environments. A model has to enter a concrete task, act, observe the result and adjust its policy from the feedback. During CoRL 2026, Chinese simulation company Guanglun Zhineng released RoboFinals RL Environments, an interactive reinforcement-learning simulation environment for robot models, as reported by QbitAI on November 17.
That supply line has become crowded in the past year. Mercor announced a $750 million Series D at a $20 billion valuation on November 10; Surge AI builds RL environments around documents, spreadsheets, web research and tool use; Handshake AI supplies expert data, RL environments and graders to frontier labs. They point at the same shift: the scarce input for post-training now includes what task to practise, where to practise it, and how to judge the result, not just algorithms and compute.
Physical environments raise the difficulty. Unlike a file or a web page, a robot's action changes an object's position, pose and contact state, and where and how hard it touches influences everything that follows. A physical-AI RL environment therefore cannot be a copy of a digital one; it has to carry real work, let the robot act repeatedly from varied starting states, and make key objects respond credibly to the policy's actions.
The first demo is an automotive repair task running more than eight minutes, combining dexterous hand manipulation, tool use, multi-stage procedure and interaction with deformable objects such as hoses, cables and gaskets. The hard part is not making any single object move, but keeping a robot interacting with objects of different physical properties across one continuous task chain.
The simulation cost is direct. Most manipulation tasks use a parallel gripper with a single open-close degree of freedom, while a dexterous hand has a dozen to more than twenty, each finger moving independently, so the action space grows sharply and multiple contact points must be solved at once. Deformable objects have to be discretised into large numbers of nodes, and small errors in stiffness, damping or friction make the simulated soft body too soft, too stiff, or prone to jitter that does not exist in reality.
The common shortcut is to split a long task into segments that each start from a hand-set initial state. That lowers difficulty but removes a key condition of real work: the robot never inherits the position, pose and contact state left by the previous step. RoboFinals links stages so the policy must keep working from imperfect intermediate states.
Underneath sits SimFoundry, Guanglun's in-house simulation base spanning solving, measurement and generation, with the generation step driven by AI agents that rebuild assets, model physics and define tasks from real jobs. The team calls the resulting standard RL Ready: every stage has an explicit completion check, physical feedback under trial and error is reliable, and the signal can go straight into training.
Calibration mixes agents with human experts. In the repair sample, a hose on an inflation device swung and bent under tension and jittered badly in the initial simulation; an agent iteratively narrowed down the Young's modulus parameter that most affected the hose response, and human experts then compared the simulated bend and swing against the real hose to make the final call. Maximum positional deviation fell by roughly three millimetres and the jitter eased.
Training and evaluation are designed as one pair: RL Envs and the RoboFinals Benchmark share task semantics, scene assets and capability definitions, while training tasks, reward tuning and evaluation instances stay strictly separate. The release also shows one task across policy iterations, from shaky motion and failed grasps to a successful screw pick at step 175. Guanglun is the only Chinese company on the Newton simulation technical committee and contributes to the SimReady standard; what to watch next is whether the environment library scales to real task distributions in home, industrial, retail and medical work, and whether model teams convert it into reproducible gains on real hardware.
Why it matters
It turns training environments into a purchasable, reusable standard component, extending physical-AI infrastructure from data and evaluation into interactive training. Whether the library scales across real task distributions and transfers to real robots is the decisive test.
Nearby Updates
All11/17, 09:28
DeepSeek rations free chat over compute shortage as V4.1 Flash tops OpenRouter at 61.2T tokens
DeepSeek is rationing its free app and website, dropping long-idle conversations and capping single inputs at about 40,000 characters because of insufficient compute. The limits land as its V4.1 Flash 1017 model leads OpenRouter with 61.2 trillion tokens in a week, though DeepSeek's own endpoint serves only 12.4% of that traffic.
11/17, 08:00
OpenAI Puts AI to Work on NATO and Allied Cyber Defenses
OpenAI has detailed its work with NATO and allied cyber agencies, focused on helping trusted defenders find vulnerabilities, test fixes and protect essential systems. The post describes a government-facing security collaboration rather than a product launch, and does not name the participating agencies, models or scale of deployment.
11/17, 08:00
OpenAI partners with Chile's government on responsible AI adoption
OpenAI has announced a partnership with the Government of Chile aimed at supporting responsible AI adoption across the country. The agreement pairs public-sector skills building with support for local startups developing AI applications, though funding and timelines were not disclosed.
11/17, 07:32
Cybersecurity AI startup Corma hits $1B valuation three months after launching its product
Corma raised a $125 million Series A led by NEA at a $1 billion valuation, about three months after introducing its product, bringing total funding to $185 million. Interest in the startup surged after a cluster of vulnerabilities and incidents on October 21 pushed security teams toward automated defense.