DribbleAMP — Adversarial Motion Priors for football dribbling
Digital Humans, ETH Zürich · team of two
Humanoid dribbling policies trained with Adversarial Motion Priors, where a learned
discriminator replaces hand-written style rewards. I built the Unitree G1 environment and the
Isaac Lab backend, and ran PPO on MimicKit with 1024 parallel environments in the Newton GPU
simulator, training on the ETH Euler cluster.
We measured what the style reward actually buys. AMP produced a natural, upright gait, but at an
equal training budget the no-AMP baseline halved speed and angle error — while dribbling in a
crouch no one would call human. Code at
github.com/ErykHalicki/DribbleAMP.
PPOAdversarial Motion Priors
Isaac LabMimicKit
NewtonUnitree G1
SLURM
Functional scene graphs with articulated object states
3D Vision, ETH Zürich · team of three
We extended functional 3D scene graphs to carry articulation state, so a robot can
track a scene as it changes instead of rescanning it. Grounding DINO, SAM and Qwen3-VL detect
objects and their parts, merged across RGB-D views.
Fusing repeated interactions cut axis-angle error by 36% (4.99° to 3.19°) and
axis-line position error by 19% (0.139 m to 0.112 m). Joint-type accuracy
rose from 86.6% to 100%. I wrote the proposal and final presentations, reviewed the team’s code,
and documented the pipeline.
3D scene graphsGrounding DINO
SAMQwen3-VLRGB-D
