A rival copied its code and ran twice as fast, and it merged the rival's changes back: how InstantTrack got 162 stars in a paper-reproduction race

A Berkeley humanoid-robot paper came out with no official code, and within a week the community spun up dozens of reproduction repos. InstantTrack showed up a week late. It didn't chase "first reproduction." It chased "fastest": one GPU, 3 minutes to train full-body motion across 1 segments on a Unitree G40. No big account retweeted it and it never hit HN. The repo got 8 stars in 162 days. The niche is small, but the approach can be copied directly for the next paper.

Nov 2: a rival said it was half as fast

On Nov 2, a developer named Fuyi Chen published a repo called BlinkMimic. The first line of its README reads "one RTX 4090, about 3 minutes," and a table below puts it head-to-head with InstantTrack v1.2 on the same card: 3.38 minutes versus 7.36 minutes to reach the target. The code wasn't written from scratch either. The README states it was modified from InstantTrack v1.2. The next day, MindPattern, an auto-aggregating AI news site, published a story with the headline naming the comparison: "...about half the time of InstantTrack v1.2."

When someone beats your code with your own code, most people get defensive. InstantTrack's author, Huy-Quang-Dao (who works at the Vietnamese robotics company VinMotion), did something else. On Nov 4 he shipped v1.3, merged all of BlinkMimic's changes back in, and the top paragraph of the README reads "v1.3 incorporates the changes from BlinkMimic... thanks to its author." He then added several GPU-level optimizations, running data collection and model updates concurrently on separate CUDA streams. Result: all 5 random seeds hit the target at the first checkpoint (3 minutes).

That back-and-forth was the two days it gained stars fastest: 11-03 +20,11-04 +33 (daily GitHub star counts, 11-06).

What it actually is

An open-source codebase for training robots to "imitate human motion." Feed it 40 segments of human motion capture data (LAFAN1: walking, running, jumping, falling and getting back up, 2.45 hours in total), and it trains a policy in simulation that has a Unitree G1 humanoid perform all 40 segments without falling once.

Its "score" sits in the GitHub description and keeps changing. The first version was 15 minutes, v1.2 was 7 minutes, and v1.3 says "about 3 minutes," more than 10.9x faster than the original paper's method (with EMA added) reproduced by the same author.

Why now: a paper with no code

Around Oct 21, the Berkeley paper SuperMimic from Pieter Abbeel's and Koushil Sreenath's groups went public (CoRL 2026 Oral). It uses a differentiable world model to train a whole-body controller on one H100 in three hours, using over a thousand times less compute than cluster-scale approaches like NVIDIA's.

The paper went viral and the code didn't ship. As of Oct 28, the community-maintained awesome-supermimic page still read "the official website has no link to training code." The gap filled up fast: ETH's Chong Zhang wrote supermimic_vibed in one hour of vibe coding, created the repo that day, and it now has 825 stars. Then came SuperMimic-Reproduced (405 stars), Uni-SuperMimic (260 stars), MorphMimic (183 stars)... awesome-supermimic lists 41 projects in total.

InstantTrack was created around Oct 29, by which point the "first to run" spot was long taken. It optimized on top of supermimic_vibed's code and cut the other side's 1.47 GPU-hours down step by step. Rivals compete on "does it match the paper," "does it work on real hardware," or "can it handle hundreds of thousands of motion segments." It competed on one thing only: speed.

Where the stars came from: aggregator pages and rivals, not big accounts

Data: daily GitHub star counts (from 10-30); release dates from the Releases page.

The first wave came from awesome-supermimic. That page is the entry point for the whole niche. It put InstantTrack in its Featured table with an animated GIF, titled "v1.2: 0/40 failures in ~7 minutes." The second wave came from the rival: BlinkMimic used it as the baseline every time, so readers clicked over to take a look.

Meanwhile the whole niche was cooling. The leader, supermimic_vibed, fell from 10-24 +188 a day to about ten or so now. InstantTrack pulled a second wave against the trend, through release after release.

What I couldn't find: no media outlet has written about it specifically, there are no HN posts, and I found no retweets from big accounts on X (I couldn't fully verify the X side).

What sets it apart: how it reports numbers

It isn't the only fast repo. It came to be seen as the trustworthy one because of how it reports its numbers:

• Every method runs on the same machine, the same simulator, and the same evaluation. Each change is a flag; turn it off and you get the original paper's config.
• 5 random seeds, and it doesn't report only the best run.
• The success bar is strict: return ≥ 0.82,40 segments 0 failures, held for 3 consecutive checkpoints. BlinkMimic later adopted the same standard outright, and its README calls it "InstantTrack's strict criterion."
• Logs for every run are in the repo, and every table can be regenerated with a script. The README has a section, "How to read these numbers," asking readers to read it before citing anything.
• It also openly reports where it got worse: v1.3's foot-contact consistency dropped from 0.845 to 0.779. The README lists the likely cause and says "we haven't tested this."

In research circles, reporting like this works better than any promotion: aggregator pages are willing to list it, and rivals are willing to use it as a baseline.

Its weak spots are also in its own README: 3 minutes is only the result on 40 training segments in simulation, with no large motion library and no real robot. The only issue is from 10-31, someone asking how to deploy "sim2real," and it has no answer yet. The author says the next project, InstantGMT (a general motion tracker for teleoperation), has a paper "in preparation."

Who it's competing with

Star counts as of 11-06.

Two players are already racing on speed, and the time is down to 3 minutes, with checkpoint intervals too close to tell apart. Squeezing further won't be noticeable to readers.

If it were you

Bottom line: Don't race on this. Learn the approach and save it for the next "paper goes viral, code doesn't ship" event.

Who it's for: Grad students and engineers in robotics, reinforcement learning, and model training who have a good GPU, can get a paper running within a week, and want to build stars, reputation, and a line on their resume.

Who it's not for: People without RL and GPU-optimization expertise. The barrier to these projects is the expertise itself. Vibe coding can produce a first reproduction, but not a 10x speedup. If you're hoping to make money from this, skip it too: this niche has no money, only attention.

How to do it:

1. Watch top-conference paper releases for the one with a big result and no official code. The window is roughly one to two weeks.
2. Someone usually produces the first reproduction within a day or two, so don't race for it. Pick one measurable dimension to claim: speed, memory, cost, running on consumer GPUs, or porting to another robot or language ecosystem.
3. Put your benchmark numbers in the GitHub description, and refresh them with every release.
4. Report numbers so tightly no one can pick them apart: same-machine comparisons, multiple seeds, public logs, and openly stated limitations.
5. When someone forks you and beats you, merge their changes generously and credit them in the README. Every comparison a rival makes is free promotion for you.

How to get users: Submit a PR to that paper's awesome list on day one (InstantTrack was added the day it was created). Give the aggregator sites and newsletters that follow up a number they can put in a headline. Ship releases often, so each version gives the aggregator page a reason to update.

When to get in early: For SuperMimic, it's already too late. There are 41 projects and the leaders are cooling, and once official code is released, attention on reproduction projects will drop further. The window for the next one is the first two weeks after the next paper comes out.

What would show I got it wrong: If InstantTrack-style "faster" repos keep gaining stars after the official code ships, speed is worth more than we assumed, and this line can still be pursued. If it stalls at two or three hundred stars along with the rest of the niche, the ceiling is the niche itself. It's still worth building, but don't expect big numbers.