SuperMimic Trains SOTA Humanoid Tracker in 3 Hours on a Single GPU

SuperMimic trained a SoTA humanoid tracking policy on a single GPU in just 3 hours, surpassing a baseline trained on 128 GPUs for 7 days. On November 10, blogger daibondalpha published a series of analyses whose core conclusion is that the efficiency comes not from more compute but from differentiable planning—running BPTT along imagined trajectories through a learned dynamics model, replacing the mainstream PPO route. The result is notable because training costs dropped by several orders of magnitude.

Confirmed

• Training results: single GPU, 3 hours, beating baseline policies (m1, m4) trained on 128 GPUs for 7 days.
• Core method: differentiable planning, performing BPTT through a learned dynamics model, backpropagating gradients along imagined trajectories (m4).
• Quantified efficiency: pathwise gradients have roughly 30x higher signal-to-noise ratio than score-function gradients, requiring about 64x fewer imagination steps than PPO for rollouts of similar models; components such as the terminal critic were also mentioned (m3).
• Key to BPTT's revival: daibondalpha noted that DreamerV2/V3 moved to REINFORCE-style gradients and the robotics community converged on PPO, so academia had largely abandoned BPTT; the variable that actually makes BPTT viable now is the control interface—50 Hz PD position control keeps gradients from exploding (m2, m5).

Why it matters

• The work shows the cost of training top-tier humanoid tracking policies can fall from a hundred GPUs for a week to a single GPU for hours, a huge deal for resource-limited research teams.
• The analysis argues that the control interface frequency (50 Hz PD position control) is the decisive factor in gradient stability—offering an alternative to the "the algorithms regressed" narrative for why BPTT was abandoned and why it works now, potentially shaping future research in differentiable planning and humanoid control.

2026-11-10 ~ 2026-11-10 · 5 related posts

• SuperMimic trains humanoid tracker in 3 hours on one GPU, beating 128-GPU SoTA baseline — daibond_alpha · 2026-11-10
• [source] Why BPTT works now: 50 Hz PD position control keeps SuperMimic gradients bounded — daibond_alpha · 2026-11-10
• [source] SuperMimic's 30× gradient SNR edge over score-function cuts imagined steps 64× vs PPO — daibond_alpha · 2026-11-10
• Why BPTT was abandoned: control interface frequency drives gradient explosion — daibond_alpha · 2026-11-10

1 near-duplicate retellings: daibond_alpha