Research
Notes from the embodied data loop.
Loop data, policy evaluation, teleop quality, and sim-to-real debugging.
2026
- Deep diveMay 12, 2026
Per-step survival: a more honest metric for VLA evaluation
A per-step survival metric for long-horizon manipulation, validated across 14 tasks and four embodiments.
- Deep diveApr 28, 2026
Closing the sim-to-real gap on long-horizon manipulation
A controlled study of which randomization axes reduce the sim-to-real gap on kitchen handover.
- Engineering noteApr 10, 2026
Real2sim with 3D Gaussian Splatting on a Franka
Rebuilding a real Franka failure in simulation with Gaussian splatting and contact replay.
- Field noteMar 22, 2026
A cross-embodiment study of OpenVLA-7B
OpenVLA-7B across Franka, UR5e, xArm-7, and ALOHA-style bimanual tasks.
- Engineering noteMar 4, 2026
When does teleop dedup actually help?
When behavioral-hash dedup helps, when it removes useful variance, and why task type matters.
- Engineering noteFeb 18, 2026
Intervention rate as a leading indicator of deployment risk
Why intervention rate can forecast deployment risk before task success moves.