HINT-WAM: hindsight self-distillation trains world action models without rewards or simulators
Phuong Tuan Dat · hf · 2026-11-19
NUS researchers propose HINT-WAM, the first on-policy self-distillation framework for post-training world action models (WAMs).
Motivation: offline SFT trains the flow-matching action expert only on noised demonstrations and never on its own sampler's intermediate states; on-policy RL struggles with sparse task rewards and credit assignment.
Method
- The student generates actions along its own denoising trajectory; at states along that trajectory the same model acts as a teacher conditioned on hindsight—the demonstration's future video—providing a dense conditional action distribution as target
- The student sees only deployment-time inputs; no environmental rewards or simulator interaction needed
- Distillation uses an energy score, a distribution-level objective matching the teacher's action distribution from the student's own samples rather than regressing the velocity field
HINT-WAM clearly beats BC control and full SFT on new LIBERO-90 tasks, RoboTwin 2.0, and other suites. Project page: tuandattt.github.io/hintwam.
More from Embodied
- GlobalFoundries to test Boston Dynamics' Atlas humanoid in chip fab in early 2027 — emmanuelvivier · 2026-11-19
- Dual-arm robot Dual Scorpion released open source, built on SO-101 hardware — rsasaki0109 · 2026-11-19
- Two weeks with Apple's always-listening Siri Recap on Apple Watch, versus Amazon's Bee — The Verge AI · 2026-11-19
- Shapeshifting Robot Aru Climbs Steel With Magnetic Feet, Already Under Industrial Contract — lukas_m_ziegler · 2026-11-19
- A nature documentary starring three flawed robots built with so-101 and LeRobot — kamathsblog · 2026-11-19
- DIY robot dog built for under $400, trained in custom MuJoCo RL sim with realistic servos — ___Mufasaa · 2026-11-19