We turn papers and published methods into working implementations. We do not stop at reproducing results in simulation; the scope extends to running and validating them on real hardware.
What you bring, what we return
| What you bring | The paper (title, authors, venue), whether code or datasets are public, the scope you want reproduced, the hardware you want to use |
| What we return | A reproduced implementation (source code), validation results on real hardware, and a report on the differences from the paper. How trained models are handled and who owns the rights are agreed before the contract. |
What we can do
- Fields: perception and recognition, grasping and manipulation, learning and control (imitation learning, reinforcement learning and others), Sim2Real / Real2Sim. The same scope as the research areas of ∂ Lab.
- Computing: training and simulation run on AWS; on-site inference and robot control run on NVIDIA DGX Spark / Jetson.
- Experimental hardware: we design hands, fixtures and mechanical parts in-house and print them on Carbon DLS and Markforged X7.
- Capturing the site: with a Leica BLK360 (accuracy ±2 cm) and a 3D scanner-type CMM, we turn the site and target objects into data for simulation (Real2Sim).
- Agreeing the scope: the hardware to be used and what is excluded from reproduction are set out in writing before work starts.
Talk to us
The paper’s title or URL, the scope you want reproduced and the hardware you want to use help the first meeting move quickly.