USC's CLAM learns robot policies from unlabeled videos, 2-3x success over baselines

ebiyik_ · x · 2026-09-30

• USC researchers presented CLAM (Continuous Latent Action Models) at IROS 2026: a scalable method for learning capable robot policies from unlabeled observation data, with paper and code open-sourced.
• The problem: learning control policies from demonstrations usually requires action-labeled expert data, which is expensive via teleoperation. CLAM targets a realistic setting where demonstrations are observation-only and only task-agnostic play data contains actions.
• Method: infer continuous latent actions between consecutive observations via self-supervised dynamics prediction, then jointly train an action decoder on a small amount of play data to ground latent actions into executable motor commands. Continuous latent actions plus joint training are shown essential for high-dimensional continuous control.
• Results: across DMControl locomotion, MetaWorld manipulation, and real-world WidowX tasks, CLAM improves average success rates 2-3x over prior latent-action baselines and approaches behavior cloning trained with privileged expert action labels.
• Takeaway: effective robot policies can be learned from unlabeled demonstrations and deployed on real hardware without expert action-labeled data.

Original post →

More from Embodied

• CrossBFM Distills a Shared Behavior Space Across Humanoid Robots in Under One GPU-Hour — Tan-Dzung Do · 2026-09-30
• Real2Gym Turns Videos Into Interactive Robot Gyms, Beating GPT-6 Direct Mode by 33% on Real Robots — Shanghai-AI-Laboratory · 2026-09-30
• Survey Maps Four Families of In-Context Learning for Robots, Toward Physical Recursive Self-Improvement — Knowin · 2026-09-30
• RobotGym launches free public MCP for Physical AI simulation with hosted GPUs — rocky_mountain12 · 2026-09-30
• NVIDIA's Robo Olympics uses Codex and natural language to train robot skills in physics simulation — RevLebaredian · 2026-09-30
• A huggable humanoid robot called Baymax shows up at IROS 2026 — CyberRobooo · 2026-09-30

Embodied channel →