One Demo, Then Practice: A Recipe for Teaching Robots New Tasks

Most robot-learning systems need a large amount of human-collected demonstrations for every new task. A new arXiv preprint, MiDAS, asks how little human data is actually necessary, and reports adapting a pre-trained generalist robot to a new task with one human demonstration and about six hours of the robot practicing on its own.

The paper introduces MiDAS as a simple recipe: take a pre-trained vision-language-action model (VLA), one of the large generalist robot models that maps camera input and instructions directly to motor commands, anchor it to a new task by imitating a single demonstration, then let the robot keep practicing on its own while a small "residual" policy learns to nudge the frozen base model in the right direction. Only that residual is trained; the underlying VLA stays fixed.

A VLA already absorbs general visual and motor skills during pre-training, so it can see and move in rough strokes without task-specific data. Fine-tuning the whole model for each new task is expensive and tends to overwrite those general skills. Training only a residual, a thin layer that adds small corrections on top of the VLA's outputs, keeps the general prior intact and concentrates the learning budget on the gap between the demonstration and reliable behavior.

MiDAS uses behavior cloning to handle the single demonstration, then value-based online reinforcement learning (RL) to handle the practice. Behavior cloning is the standard imitation move: the robot tries to copy what the human did. Value-based online RL is the practice move: the robot tries actions, gets a reward when it succeeds, and updates its residual so good actions become more likely. The result is an offline-to-online recipe: learn from one human clip, then keep learning on the robot.

The authors evaluate MiDAS on three testbeds. On the LIBERO and RoboCasa benchmarks, standard simulation suites used to compare robot-learning methods, MiDAS recovers strong task performance from as little as one demonstration and substantially outperforms baselines, including on conditions the demonstration did not cover. On a bimanual YAM robot, a two-armed hardware platform the team uses to test methods outside simulation, starting from a policy that was fragile after a single demonstration, MiDAS improved robustness and learned new successful behaviors over roughly six hours of autonomous interaction.

The authors call this, to their knowledge, the first demonstration of reliable robot policy adaptation from a single task demonstration. The scope matters: "reliable" is bounded to the LIBERO and RoboCasa benchmarks plus the YAM hardware, not a claim about general household robots, and the comparison is against other single-demonstration baselines, not against systems trained on thousands of demos.

Fully autonomous robot learning from scratch remains hard because sparse rewards, a one or zero signal at the end of a long action sequence, give the robot almost nothing to learn from during random exploration, so it cannot discover good behavior on its own. MiDAS does not solve that problem. It sidesteps it by starting from a pre-trained VLA and a single demonstration, which puts the robot near a workable policy before practice begins. The paper presents minimal-data adaptation as the closest tractable proxy for fully autonomous improvement, not as a replacement for it.

The cost has moved. The expensive line item is no longer the human data collection for every new task; it is the roughly six hours of robot practice time, on real hardware, in the loop. That is a different bottleneck, with a different budget, and a different set of failure modes: robot wear, safety constraints, and the fact that six hours of practice is still six hours, not a slogan about instant learning.

Sparse-reward, from-scratch exploration is still hard, and the single-demonstration starting point is a workaround, not a solution. The next reader-facing question is whether a residual on a frozen VLA can absorb more of that gap, or whether the field still needs a separate mechanism for robots that start with no demonstration at all.