Robotics and RL Researcher (Research Coding Tasks)

What you'll do

You start from a recent robotics and reinforcement learning paper that has released code, ideally your own. You remove the code that implements the paper's core contribution and package everything else as a task in which an AI agent has to rebuild that contribution and beat the authors' published results on the paper's own benchmarks. You write the task statement, define held-out evaluation workloads with a pass threshold for each metric, and supply a hidden reference solution plus a second solution that shows the bar can be beaten. Later you review agent attempts on your task.

Here the paper's headline result is usually a return, success rate or tracking error in simulation, and the ad names robot learning, sim-to-real transfer, control and reinforcement learning as the areas of depth. Held-out evaluation would probably mean unseen environments, seeds, tasks or perturbations, with a threshold per metric.

It is remote, asynchronous and task-based.

Who fits

• Published research in robotics and reinforcement learning with released code, at venues such as CoRL, RSS, ICRA, IROS or NeurIPS, or comparable
• Released code for at least one of your papers (a GitHub or artifact link)
• A PhD in computer science or a related field, completed or in progress
• Preferred: first-author papers that report measured improvements over a baseline; comfort with Docker and reproducible benchmark setups; depth in the areas listed above

The application form asks for a link to a paper you authored and its code, and the headline result from that paper; both are required. AfterQuery says candidates who meet the requirements and supply the materials are prioritised, and that it runs thorough background checks, so apply only if you meet the bar.

What it pays

The board headlines $80/hour. The ad's own text describes the pay differently: $200 per approved task, with each task taking a few hours. The two figures line up only if a task takes about two and a half hours, so treat $80/hour as a nominal rate and the per-task figure as what is actually paid. Payment is per approved task, so a rejected task earns nothing. The ad does not publish payment cadence or method; see the company page.

Worth knowing

Good:

• Your own research and code are the raw material, so little time goes on learning an unfamiliar codebase
• Fully remote, asynchronous and flexible
• A stated per-task figure rather than an open-ended estimate

Less good:

• Pay is per approved task, so the quality of your task design decides what you earn
• Writing a fair held-out benchmark and two working solutions is the hard part, and the ad estimates only "a few hours"
• RL results are noisy, so the threshold has to be set against variance across seeds, and a bar the authors' own method clears only sometimes makes for a bad task. The ad does not say how sim-to-real claims would be checked; expect them to be reduced to simulated proxies, and ask before assuming otherwise. NeurIPS appears in the venue list, so ML-first RL researchers are in scope as well as roboticists.
• No hours commitment or task volume is stated, so steady income is not guaranteed

About this listing

Read on the AfterQuery Experts board in October 2026, where it was posted on 29 October 2026. It is one of a dozen research coding task roles that share a template, each tied to a different field; the pay, process and application form are the same across them. See AfterQuery.