120ftMaster's Thesis: Robot Learning from Operational Data
Stockholm · Internship
Why join 120ft?
Picture products built by robots in shipping containers beside the customers who buy them, and 120ft, a Stockholm industrial robotics startup, could be the company that makes that routine. Four angel backers have put $646K in, which is modest, but nine open roles, six in engineering, show the build-out has started. Two of those roles opened in the last 30 days, so early engineers will set the pace.
JustBacked's take, based on 120ft's own site and its funding and hiring data.
Can robots learn from the work they already do? In this master's thesis at 120ft, you will investigate how operational data can improve robot learning for real assembly tasks.
Host: 120ft Factory AB, Stockholm
Company supervisor: Sulthan Suresh Fazeela
Field: Machine learning and robotics
Background
120ft builds Factory OS, a software layer that lets humanoid robots, currently the Unitree G1-D, carry out high-mix, low-volume assembly work. Robot behaviour comes from Vision-Language-Action (VLA) policies trained on recorded demonstrations.
Today, almost all training data comes from dedicated teleoperation sessions. Every hour of that data costs an operator's hour. But the robot also generates data every time it runs.
The research question
Operational rollouts include per-frame tags showing whether a human or a policy was driving. You will investigate whether data from policy-driven episodes can improve the policy, how it can be combined with teleoperation demonstrations, and which data helps rather than hurts.
What you will investigate
This is an offline, experimental project using already-collected rollout data, rather than building a live production pipeline. The goal is to characterise the data and give an evidence-based recommendation on its value for training.
Compare the operational dataset with curated teleoperation datasets: volume, success and failure balance, and recurring failure modes.
Test whether fine-tuning on operational data changes policy quality, and whether filtering or weighting by operator-judged outcome labels makes a difference.
Explore other data-centric questions that emerge, including the reliability and consistency of outcome labels.
What we provide
A technical supervisor.
Access to existing operational and curated datasets.
Access to the training stack for fine-tuning experiments.
Access to a real Unitree G1-D if additional rollouts are needed.
Who this is for
You are a master's student in machine learning or robot learning, comfortable with empirical experimentation and data analysis.
Apply
Send your CV and a GitHub, LinkedIn or website link. Tell us why this thesis interests you and share an example of work you are proud of.