Astribot Releases SmoothRL for Asynchronous Robot Learning; Qujing and Moore Threads Announce AI Token Deal

According to Astribot's paper, the problem is that when inference runs asynchronously, a robot may not execute the whole action chunk: some actions are committed to the current execution, others are discarded when the next chunk arrives. Training on the entire chunk can reward or penalize actions that never occurred. SmoothRL divides chunks into committed, execution and discarded regions and updates the policy only through the execution region, i.e., the actions really performed. The authors call this “Reinforce in Deployment” because training rollouts use asynchronous inference just like deployment.

The team fine-tuned π0.5 as the base policy for each task and added a lightweight TD3-style residual policy in raw action space. On Astribot's S1 robot, action execution ran at 30 Hz and inference was requested at 5 Hz; each action chunk had 32 frames, 20 of which were overwritten by subsequent chunks before execution.

In real-robot trials, success rates rose from 39% to 94% for dynamic throwing, 8% to 83% for pen capping, and 30% to 90% for parcel opening. The parcel-opening curve was not steadily improving: it dipped from 30% to 20% around 150 rollouts before reaching 90%. Adding a smoothness constraint lowered the right end-effector's root-mean-square acceleration by 52% and jerk by 47% in an autonomous throwing rollout. After training, failed samples clustered near correct positions, indicating systematic bias had been reduced.

Astribot notes the current implementation assumes each chunk inference finishes within a preset latency budget, and the frozen base policy bounds what the residual policy can correct; if the base policy is too far from the target behavior, local corrections cannot fix it. Planned follow-ups include larger policy updates, wider task distributions and end-to-end optimization between asynchronous execution and generative policies. Wang Jianan led the project; Gao Guang and Nong Yuxuan were co-first contributors, with Huang Baifu as participant.

On the compute side, Qujing Technology and Moore Threads said the September 3 agreement integrates Qujing's domestic PD heterogeneous inference technology with Moore Threads' MTT S5000 cards and MUSA software, and is meant to create domestic high-quality AI token factories around Qujing's ATaaS platform. The combined solution has entered production and is carrying real traffic from a leading model vendor, they said. By assigning MTT S5000 to the prefill phase and reserving high-bandwidth GPUs for the decode phase, the two companies say the system outperforms an international advanced computing solution in cost per high-quality AI token under the same production-service standards.

The companies reported production measurements of average token generation speed above 50 TPS, a KV cache hit rate above 90%, stability of 99.9%, and production-grade low time-to-first-token. Four to five MTT S5000 cards used as a prefill pool were said to give better input-token price-performance than international reference systems. They plan to deliver the results as a Token Pod—a standardized super-node combining the heterogeneous engine, ATaaS platform and operational services. Qujing says it had co-built projects at a daily trillion-token scale and had built daily hundred-billion-token capacity in multiple projects by August 2026. The partners will extend the approach to internet companies, frontier foundation-model firms and operators, expanding what Qujing calls its “fewer models, deeper optimization” route.