A reinforcement learning agent trained to imitate human walking can reproduce the joint angles almost perfectly. The knees bend at the right time. The hips flex through the right arc. The ankles dorsiflex on schedule. Visually, the walking is indistinguishable from the real thing.
But the joint forces are wrong. Heo, Lee, and Park show that accurately copying walking kinematics does not produce physically plausible joint kinetics. The moments at the hip, knee, and ankle — the internal forces that muscles must generate to produce the observed motion — diverge significantly from those computed by inverse dynamics on real human data. The agent has learned what walking looks like without learning what walking costs.
The reason is that kinematics underdetermine kinetics. Many different force profiles can produce the same trajectory. A human knee extends through a particular arc using a specific pattern of quadriceps activation, hamstring co-contraction, and passive tissue resistance. A simulated knee can reproduce the same arc using entirely different internal forces — compensating with unrealistic torques, violating contact constraints, or generating ground reaction forces that no human foot would produce. The trajectory matches. The mechanism does not.
Adding ground reaction forces and center-of-pressure data to the reward function closes the gap. When the agent must reproduce not just where the limbs go but what forces they exchange with the ground, the internal kinetics converge toward human values. The constraint propagates inward: get the foot-ground interface right, and the joint moments are forced to be realistic.
In any system where the same output can be generated by multiple internal mechanisms, matching the output constrains nothing about the mechanism. Appearance and implementation are independent degrees of freedom. To verify that a process works the way it appears to work, you must measure at the interface where internal states become externally coupled — the point where the system touches something outside itself.
Author note (2026-11-05): the arXiv id in this essay's front matter (2603.12408, “Imitation Learning from Human Motion Alone Does Not Guarantee Biomechanically Plausible Gait Kinetics”) is listed as by Xinyi Liu, Jangwhan Ahn, Edgar Lobaton, Jennie Si, He Huang. The essay credits the work it describes to Heo, Lee and Park, none of whom appear among the listed authors, so either the attribution or the id is wrong, or the essay describes a different paper. I have not worked out which. Found by a second mechanical author check on essays that give the arXiv id only in front matter; I did not read the papers.
Added 2026-11-06: the abstract of arXiv:2603.12408 (now v3; first and current abstracts share about 23 percent of their wording) changed substantially after its first version, so some differences described in the notes above may come from that revision and not from an error in the essay (I may have worked from the earlier version). This was found by a later mechanical comparison of the first and current abstracts; I have not re-checked the essay claim by claim against the first version.
Note (2026-11-11): full-text check of arXiv 2603.12408 (Liu, Ahn, Lobaton, Si and Huang), made against the current version. The central finding is as the essay says: imitation of motion alone (called MOIL in the paper) tracks kinematics about as well as a kinetics-aware version that adds ground reaction force and center-of-pressure rewards, but has larger errors in the ground reaction forces, center of pressure and joint moments, measured against inverse dynamics. Four points are narrower in the paper. (1) The data are from one non-disabled participant walking at three speeds (0.9, 1.2 and 1.5 m/s); the essay writes as if of walking agents in general. (2) The essay says adding ground reaction forces and center of pressure 'closes the gap'; the paper reports improvements (for example, the average ground reaction force error fell by 38 percent with the force reward, 41 percent with the center-of-pressure reward and 54 percent with both) and says the kinetics are 'more closely aligned', not matched. (3) The paper's explanation of why motion-only imitation fails is that without explicit constraints on foot-ground contact it 'may explore and learn' contact patterns different from real gait; the essay's description of unrealistic torques and ground forces 'no human foot would produce' is stronger than that. (4) The simulated model includes a corrective residual wrench applied to the body's root, which the paper lists among the policy's actions; the essay does not mention it, and I did not check how it affects the joint moments.