TWIST2
Honest P10 option: reproduce in sim, deploy the manipulation half on the real SO-ARM101.
Read on arXiv ↗
The papers that make this possible, from PPO to TWIST2.
Honest P10 option: reproduce in sim, deploy the manipulation half on the real SO-ARM101.
Read on arXiv ↗The predecessor; full whole-body control via mocap. The “why whole-body” argument.
Read on arXiv ↗Universal human-to-humanoid whole-body teleop and learning.
Read on arXiv ↗Humanoid shadowing & imitation from human video.
Read on arXiv ↗Loco-manipulation via an isomorphic exoskeleton cockpit (decoupled-control baseline).
Read on arXiv ↗Closed-loop whole-body teleop for long-horizon tasks (partial-control baseline).
Read on arXiv ↗Teleoperation with immersive active visual feedback.
Read on arXiv ↗Predictive motion priors for humanoid whole-body control.
Read on arXiv ↗Optional P10 retargeting study: human→robot on CPU.
Read on arXiv ↗The core policy-gradient algorithm behind Isaac Lab / RSL-RL.
Read on arXiv ↗Massively-parallel RL for legged locomotion (RSL-RL foundations).
Read on arXiv ↗Action-chunking transformer; LeRobot's recommended first policy for P4.
Read on arXiv ↗Visuomotor policy via action diffusion; the P4 comparison policy.
Read on arXiv ↗Generalizable visuomotor policy from simple 3D representations.
Read on arXiv ↗The lineage behind TWIST2's high-level policy.
Read on arXiv ↗Pretrained visual representations for manipulation (backbone).
Read on arXiv ↗Vision-language-action flow model; manipulation-at-scale reference for P10.
Read on arXiv ↗VLA with open-world generalization.
Read on arXiv ↗Real2sim2real; contextual humanoid control from video.
Read on arXiv ↗