The authors designed SEAR, a unified framework that runs an LM agent in three modes: code-as-policy (writes a control program), agent-as-policy (acts live in the loop), and auto-research (fine-tunes a VLA on self-collected rollouts).
Benchmark reveals when robotic agents truly learn and when they just retry
SEAR framework tests seven language models across 330 simulated and five physical tasks, finding that learning curves depend on task difficulty and budget, and code skills transfer more reliably than task-specific solutions.
Industry
Xinyi Wang · Bangzheng Li · Hanyu Wang · Lin Zhang · Andrea Yaoyun Cui · Zhenggang Tang · +2 more
Luma AI
Research Digest··3 min read
Wang et al.
Why this paper
From Luma AI
In one line
Self-improving robot agents do not necessarily learn from experience; apparent gains can come from retries or luck, and learned skills do not always transfer.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§