KAIST's FRED Self-Distills Single-Step Video Representations, Boosting Robot WAMs on Real Tasks
coallaoh · x · 2026-10-24
Researchers at KAIST and RLWRLD introduce FRED, a framework for World Action Models (WAMs) that derive actions from future video representations extracted in a single denoising step. FRED improves representation quality via self-distillation—using representations from ground-truth future observations as targets—while conditioning on observation history for temporal context. Across real-world manipulation scenarios FRED consistently outperforms competing WAMs, with the largest gains on long-horizon and out-of-domain zero-shot tasks. Human-to-robot skill transfer improves substantially, reaching 50.0% success on tasks absent from robot training data. Paper, code and checkpoints are released.
Related event: KAIST's FRED Self-Distillation Boosts Robot Task Success from 19.4% to 50%(2 posts)→
More from Embodied
- NVIDIA team recruiting 2027 interns for humanoids, motion generation, and RL research — xbpeng4 · 2026-10-24
- MIT's blind assembly robot hits 73% peg-insertion success with zero-shot sim-to-real — chris_j_paxton · 2026-10-24
- "RL doesn't work until it does": researchers showcase RL for high-reliability robot manipulation — chris_j_paxton · 2026-10-24
- mujoco_warp_pbr: batched ray-traced PBR renderer for MuJoCo in NVIDIA Warp — mishig25 · 2026-10-24
- Claude Code taught a Roborock to escape from under the stairs, 10/10 success — fillyourcoffee · 2026-10-24
- NVIDIA expands Isaac robotics platform at GTC Berlin; enact_intel stealths with $30M pre-seed — lukas_m_ziegler · 2026-10-24