Tianqi Zhang

CS Ph.D. student @UNC | B.S. @Tsinghua_Uni

[Jiatao Gu @thoma_gu]

Video benchmarks aren't saturated, they're just easy to game. We introduce Video-Index by dropping all shortcut-solvable questions across 115 benchmarks — key findings: Astra 79.3% beats humans 55%, while open models <20%. Led by my incredible first-year PhD student @EnxinSong !

[Yikai Wang @yk20wang]

Introducing APEX — a unified policy for adaptive humanoid traversal across high platforms up to 80 cm (114% leg length).

- Without reference motion, a generalized ratchet progress reward formulation enables learning of a set of adaptive, contact-rich full-body maneuvers.

- Through distillation, we obtain a single context-aware policy that integrates these maneuvers with cyclic locomotion.

- Deployment relies on elevation mapping from a single LiDAR.

The system demonstrates robust adaptation, autonomous skill selection, and smooth multi-skill transitions.

Website: https://t.co/zxLhia5yVQ
Paper: https://t.co/jtYVE6r7WX

Huge thanks to my co-authors:@lengtx20, @changyi_lin1, @shiqiliu_67, Shir Simon, Bingqing Chen, Jonathan Francis, @zhao__ding. Also grateful to @yxyang1995 for insightful discussions and valuable feedback.

[Yikai Wang @yk20wang]

Check out KPI: a general real-time interaction kernel on top of a humanoid whole-body tracker, unlocking diverse contact-rich loco-manipulation while preserving general-purpose tracking. Pair it with a powerful agent like GPT-6 Astra to turn high-level instructions into physical interaction!

[Tianqi Zhang @AlanZhang0928]

A robot can follow a reasonable interpretation of an instruction and still fail the benchmark.

Meet BenchMend: we use GPT-6-Astra to uncover instruction–verifier gaps in robotics benchmarks, then repair the wording while preserving success checkers and recorded demonstrations.

Key takeaways:

📝 Instructions matter. Small edits can substantially improve alignment between task language and benchmark verdicts.
🤖 Even strong agents have limits when performing zero-shot robot control. Tasks requiring precise contact, placement, and coordination remain challenging for agents.
🛠️ An imperfect policy can still help debug benchmarks. Agents that diagnose gaps, propose fixes, and test them could enable self-improving benchmarks and potentially support recursive self-improvement (RSI).

Read the blog: https://t.co/nbYo5rOgtU
Watch the rollouts: https://t.co/Lk8axNRfpN