Experiment-free learning of exoskeleton assistance is not an unsolved problem

We present three quantitative methods: 1) estimation of exoskeleton mechanical power and energy ratio from published data, 2) a systematic review of the exoskeleton literature on reported energy ratios, and 3) timing correction analysis of the replication experiment, to address concerns raised by Collins et al. (2026) about Luo et al. (2024). Together, these analyses support the reported metabolic reductions and the validity of exoskeleton control via learning in simulation. The critique rests on an unsupported premise: that exoskeleton energy ratios above 4 are physiologically implausible. This premise of Collins et al. (2026) is not supported by the cited evidence, and the error originates in their own cited source. Sawicki and Ferris (2009), the paper they invoke as authority for the limit of 4, state explicitly that "reported values of the muscular efficiency range from 0.10 to 0.34, with many sources assuming an average of [~]0.25." The value of 4 corresponds to this average, it is not a physiological ceiling. Treating an average as a physiological upper limit is a fundamental error. The published exoskeleton literature further contradicts the claim, including work by the authors of the critique themselves (Collins et al., 2015: 4.3; Young et al., 2017: 5.0) and independent work (Malcolm et al., 2013: 4.8; Seo et al., 2017: 6.7). In contrast, our walking energy ratio is 2.4, calculated directly from Fig. 4 of our paper. Our device delivers higher peak torque (14.1 Nm vs. 10.9 Nm, Lim et al., 2019) and achieves a slightly larger metabolic reduction (24.3% vs. 21%). Independent groups have since demonstrated meaningful metabolic reductions using learning-in-simulation frameworks, including Barati et al. (2026, 15.2% mean and 22.5% maximum) and Zhou et al. (2025, [~]20% during running). The claim of Collins et al. (2026) that this problem "remains unsolved" is directly contradicted by these independent results. The experiment in the critique is not a valid replication of our method. Our controller is a neural network with [~]10,000 parameters learned through deep reinforcement learning in musculoskeletal simulation; the critique instead applies a pre-programmed fixed torque curve with no learnable parameters. Beyond this, the replication contains three methodological errors: 1) a heel-strike timing assumption producing offsets up to 30% of the gait cycle; 2) an averaged torque profile that discards subject-specific control; and 3) a device [~]50% heavier than ours (4.8 kg vs. 3.2 kg) without measuring the metabolic penalty of the added weight. The critique also misreports Samsung data, with reported values approximately double those in the original publication, errors that directly underpin their physiological limit argument.