MiDAS lets robots learn new skills from a single demonstration
One demo and six hours of practice is all MiDAS needs to master new robotic tasks.
The paper, titled 'Adaptation of Generalist Robot Policies with Minimal Data,' tackles one of robotics' hardest problems: reducing dependence on massive, task-specific human data. The authors propose MiDAS, a two-stage offline-to-online reinforcement learning (RL) recipe. First, a pre-trained generalist VLA policy is anchored to the target task using behavior cloning on as little as one demonstration. Second, value-based online RL improves the policy via a residual parameterization, letting the robot autonomously explore and refine its behavior beyond the initial demonstration.
Across benchmark suites LIBERO and RoboCasa, MiDAS reliably recovered strong success rates from a single demonstration, substantially beating existing baselines and even generalizing to situations not seen in the demonstrations. The team also validated MiDAS on a real bimanual YAM platform, showing that a fragile policy built from one demo could be transformed into a robust one capable of discovering new successful behaviors after roughly six hours of online interaction. This work—from researchers including Max Simchowitz and Aviral Kumar—is notable as the first demonstration of reliable robot policy adaptation from just a single task demonstration, moving toward robots that can improve autonomously with minimal human guidance.
- MiDAS uses behavior cloning on one/few demonstrations to anchor a pre-trained VLA, then applies value-based online RL on a residual policy.
- Outperformed baselines on LIBERO and RoboCasa benchmarks, generalizing beyond demonstrated conditions.
- First reliable single-demo adaptation, validated on a real bimanual YAM robot in ~6 hours of online interaction.
Why It Matters
MiDAS could slash the cost of teaching robots new tasks, enabling rapid deployment on new jobs with a single human demo.