Phoenix Tech Report Tackles Why Robots Fail at Final Contact
jiqizhixin · x · 2026-11-03
Anyverse Dynamics and CASIA's RL team released Phoenix, a technical report on why robots follow plausible trajectories yet fail at final contact — geometry alone doesn't guarantee useful tool contact.
Key ideas:
- Dual-teacher distillation into a DINOv3-initialized Phoenix visual encoder (geometry + temporal teachers)
- Phoenix-ILA learns interaction latents from synchronized motion and tactile sequences
- CS-VAE for causal compression; Phoenix-MWA uses a MoT with video/action experts and joint attention to model future visual states and actions
- PhysCanvas provides an action-conditioned visual-tactile simulator for pre-deployment policy refinement
- Supports 18 arm groups (Franka, UR, ARX, ALOHA, etc.) for cross-embodiment transfer
Reports improved real-robot success on test-tube insertion across three tested arms, avoiding treating every real-world failure as a new training experiment.
More from Embodied
- YC-backed Cyclon ships AI phone OS on Pixel 8a, promising apps that act for you — Scobleizer · 2026-11-03
- YC-backed Subvocal builds RF smart glasses that decode silent speech into text — aarushyadav · 2026-11-03
- GPT-6 Astra tops robot index, but finishes dual-arm sim task only twice in 800 trials — minchoi · 2026-11-03
- Starcloud-1 marks one year in orbit with the only NVIDIA H100 in space, 3 more launches planned — ycombinator · 2026-11-03
- Ant Group's Robbyant shows LingBot-VLA 2.0: 60k hours of real data, runs on 20+ robots — huangyun_122 · 2026-11-03
- Ropedia W0: cross-embodiment world model trained on 10k hours of human video hits SOTA — liuziwei7 · 2026-11-03