University of Michigan researchers report improved humanoid walking-trial success rate with PRISM

humanoid.guide reported that University of Michigan researchers conducted simulated humanoid walking trials using PRISM. The share of 24-second trials completed without a fall was 92.5%, higher than the 51% for a conventional MLP; an MLP with about 43% more parameters scored 52.3%.

In the reported comparison, the MLP with about 43% more parameters had a completion rate of 52.3%, not much higher than the conventional MLP's 51%, while PRISM scored 92.5%. Although the results are from simulation, the comparison underscores the need to evaluate walking-control performance in terms of input design as well as network size.

The comparison focuses on network size and how inputs are handled. PRISM had a completion rate of 92.5%, compared with 51% for a conventional MLP. The finding that an MLP with about 43% more parameters still scored only 52.3% suggests that, at least under the reported simulation conditions, simply scaling up the network does not readily explain the difference. However, this comparison alone does not establish that PRISM's approach is generally superior to other control methods. According to the article, PRISM's architecture handles interactions among observations, past actions and task commands at the input stage. When evaluating control results, it is necessary to look not only at the final success rate but also at the conditions under which differences emerged. The article says the advantage persisted under conditions in which friction, mass and payload differed from those used in training, but the results presented here are limited to simulation. Results showing performance under changing conditions should therefore be distinguished from evidence of fall prevention or stable walking on a physical robot. Comparisons of walking-control methods require evaluations that align network size, input representation and trial results under the same conditions. The figures in this report invite comparisons among methods, but the provided article text does not specify the total number of trials, the simulation environment or how a fall was defined. No official University of Michigan announcement or paper was provided as a source, so readers should consult the original sources for details. No related articles from this newspaper were provided, so no comparison with past coverage is included. Three points warrant further scrutiny: first, the experimental conditions, including the number of trials and environment settings; second, the difference between results on a physical robot and in simulation; and third, whether the performance gap is replicated under conditions beyond changes in friction, mass and payload.