The Pulse
Robot Policy Readout Nearly Halves Xiaomi Inference Latency
A paper reports that Gaussian Action Readout reduced measured batch-one inference latency for Xiaomi-Robotics-1 by about 48%, with task success close to the original policy in simulation. The readouts take about 115 seconds to fit once feat
AI.info Team · Written with AI · fact-checked ·
A robot policy can keep nearly the same task success while replacing several rounds of action generation with one network evaluation, according to a paper posted to arXiv on October 29. The method, called Gaussian Action Readout (GAR), cuts measured batch-one inference latency by about 48% for Xiaomi-Robotics-1 and 36% for NVIDIA’s GR00T N1.5 in the researchers’ tests.
Youssef Saied, Ziyad Sheebaelhamd, Claire Vernade and François Fleuret propose fitting small readout layers to a pretrained policy’s frozen internal features.
“It samples with one action-network pass, gives closed-form likelihoods, and achieves near-teacher task success at up to twice the inference speed.”— the authors write in “Gaussian Distillation of Flow Policies”. The results come from simulated manipulation benchmarks; the paper does not report a physical-robot evaluation of GAR.
One evaluation replaces four or five flow steps
Flow-based robot policies produce actions by repeatedly evaluating a network as it transforms noise into an action chunk. That iterative process can model varied possible actions, but it adds computation at execution time. GAR instead predicts a Gaussian distribution for the continuous actions, plus probabilities for binary commands, from the policy’s existing readout features.
The method keeps the pretrained teacher networks frozen and trains only small readouts. For its main RoboCasa365 experiments, the team fits them using 6,000 contexts, equivalent to 0.02% of eligible training states. The contexts span 300 tasks; for each task, the fitting procedure draws four interior states from each of five randomly selected human demonstrations.
At deployment, a single action-transformer evaluation supplies the Gaussian parameters, and the policy samples directly without integrating the flow field. That reduces five action-transformer evaluations to one for Xiaomi-Robotics-1 and four to one for GR00T N1.5. Once features and teacher targets are cached, fitting Xiaomi’s 6,000-context diagonal readouts takes about 115 seconds. The paper reports preparation costs separately: feature extraction and teacher sampling used 3.84 GPU-hours in total, with the longest parallel job lasting 1.14 hours. Download, extraction and data checks took 2.71 hours of wall time; motion and command target estimation took 6.9 seconds.
Latency falls, while task success stays close
The researchers measure performance on 50 tasks selected from RoboCasa365, a simulation benchmark of household manipulation. They run 50 episodes per policy on each task, for 2,500 episodes per policy. Xiaomi-Robotics-1’s overall success rate is 58.24%, compared with 57.64% for its diagonal GAR version; GR00T N1.5 scores 23.08%, compared with 21.92% for GAR.
Measured batch-one latency, including vision-language processing, falls from 220.25 to 115.36 milliseconds for Xiaomi-Robotics-1, a 1.91-times speedup. For GR00T N1.5, it drops from 68.01 to 43.30 milliseconds, a 1.57-times speedup. Because vision-language processing remains part of the timed pipeline, it limits the total gain even though GAR removes multiple action-transformer evaluations.
The results vary by task group and model. Xiaomi’s diagonal GAR scores 80.56% on seen atomic tasks and 31.50% on unseen composite tasks, against teacher results of 81.11% and 32.00%, respectively. GR00T’s teacher-to-GAR scores are 46.00% and 45.78% on seen atomic tasks, but only 5.00% and 5.25% on unseen composite tasks; the paper cautions that the small gap there is less informative because teacher success is already low.
The Gaussian shortcut has limits
GAR relies on the paper’s finding that the continuous-action distributions of the studied flow policies are well approximated by Gaussians. The authors test that assumption against teacher-generated actions and report similar task success for Gaussian sampling and teacher sampling in their controls. Their approach is not a general replacement for any robot policy: it reuses frozen features from specific pretrained flow-policy teachers, and its reported task results come from simulation.
The team also evaluates a separate Xiaomi-Robotics-0 teacher on the ten LIBERO-Long tasks. Diagonal GAR reaches 95.4% success, compared with 96.6% for the teacher and 96.0% for the teacher’s raw one-step execution. In another simulation test, Xiaomi’s GAR policy produces two different successful object-moving orders from the same starting context, though only three of 32 attempts succeed on that task.
Those findings point to the paper’s practical trade-off: GAR removes repeated flow steps without making the tested policies deterministic, but its strongest reported result is lower latency with near-teacher performance in simulation—not proof of faster or equally capable control on physical robots. The authors say code will be released soon.