Astribot's SmoothRL gives robots online learning that survives a thrown ball
Online reinforcement learning for real robots usually assumes the world will politely pause. Astribot's new framework shows it doesn't have to — and the gains come from correcting the actions that already started, not the ones the model just planned.
Stardust Intelligence (Astribot) released SmoothRL, an asynchronous online reinforcement learning framework built for high-dynamic manipulation. The lead author is Astribot's Wang Jianan, with Gao Guang and Nong Yuxuan as co-first authors; the research note and technical report are published on the company's research page. The framework is the first asynchronous online RL system validated on real, high-speed robotic tasks, and the headline result is an end-to-end throwing motion whose end-effector acceleration RMS drops 52 percent and jerk drops 47 percent after on-policy correction — meaning the robot accelerates and decelerates less abruptly mid-throw, with smoother trajectories and fewer stalls.
The core problem SmoothRL attacks is mismatch between planning and execution. Vision-language-action (VLA) and world-action models typically emit a chunk of future actions in one inference call, and inference time on large models can stretch to hundreds of milliseconds. A robot mid-throw cannot stop and wait. Traditional online RL assumes the policy generates, the robot executes, and the system updates — a clean one-to-one mapping. Asynchronous execution breaks that correspondence: a new chunk arrives before the old one finishes, and only part of the planned action actually lands in the physical world. SmoothRL partitions each action chunk into three execution states — what the robot already committed to, what it actually executed, and what was preempted — and updates only the parts that the body really felt. That lets the system absorb non-blocking inference and still get a clean gradient signal.
On the Astribot suite the framework cut systematic bias — release timing in throws, lateral drift in box-opening, repeatable placement errors when capping pens — and it did so with sparse binary success-or-failure rewards. Operators can still intervene mid-rollout and have those corrections flow into training, which makes this a deployable online post-training pipeline rather than a closed-loop lab demo. The current implementation requires every chunk-level inference to fit a preset latency budget, and the residual correction policy is bounded by how good the frozen base policy already is. The team says the next step is wider policy updates and end-to-end optimization that couples the asynchronous execution loop back into the generative policy itself.
Moore Threads and Quxing sign on to a Token Pod
Moore Threads and Quxing Technology signed a strategic partnership that turns the MTT S5000 domestic AI accelerator into a production prefill card for a jointly branded Token Pod. The agreement was announced September 3 in Beijing, with Quxing's PD-heterogeneous (prefill-decode split) inference stack sitting on top of Moore Threads' MUSA software platform. Quxing chief scientist Wu Yongwei, founder Ai Zhiyuan, CFO Wu Wenjie, and chief architect Xie Weiyu signed for Quxing; Moore Threads CEO Zhang Jianzhong and senior VPs Dong Longfei and Luo Wenyong signed for the chipmaker.
The pair are pushing a Token Pod as a standardized production unit for high-quality AI Token generation — a rack-scale combination of S5000 prefill cards handling prompt ingest and KV-cache build, paired with high-bandwidth GPUs running the decode side, all scheduled by Quxing's ATaaS platform. The pitch is that the cost edge is not from raw silicon pricing but from treating prefill and decode as independently pooled, billed, and optimized resources, then folding the resulting configuration, traffic, and scheduling policies back into ATaaS as reusable production templates. After several rounds of joint tuning, MTT S5000 has cleared model-compatibility testing and is now serving real production traffic rather than running as a benchmark-only card.
The broader play is that domestic chip + domestic inference platform combos can now ship as repeatable deliveries to internet companies, frontier model labs, and telcos — the kinds of buyers who care about both unit cost and supply-chain durability. Quxing is already standing up a 10,000-card-scale AI Token factory in Hangzhou's Qianjiang Century City as part of its eastern-China build-out, and the Token Pod abstraction is what makes that scale legible to procurement teams that want a unit of capacity, not a stack of vendor relationships.
What to watch: whether SmoothRL's three-zone action accounting gets adopted by other VLA training stacks — and whether the first async-RL commercial deployment outside Astribot lands before the end of the year.
Online RL for robots is finally leaving the lab — but only if the world agrees to wait. Is async execution the missing layer, or will the next bottleneck be the frozen base policy itself? Tell us in the comments.
Sources: QbitAI — SmoothRL launch · Astribot research — SmoothRL technical report · QbitAI — Quxing + Moore Threads partnership
