This seems to be an interesting new Chinese paper on robotics!
Did these researchers successfully overcome catastrophic learning of previously learned skills?
"... Researchers ... recently developed HOST (Human-to-robot One-Shot Skill AcquisiTion), a new framework that could allow robots to acquire new skills faster and more efficiently. Their proposed learning approach ... allows a robot to acquire a new skill from a single video showing a human demonstration without compromising previously acquired abilities. ..."
"...
Highlights
- Inference-time skill acquisition: one human video, no fine-tuning, and no parameter update.
- Fast acquisition: 29 seconds per novel skill on average, including recording the demonstration.
- Broad real-robot evaluation: acquires executable skills across 50 novel manipulation tasks, each evaluated over 20 trials.
- Strong novel-task performance: achieves 62% average success on the task subset used for baseline comparisons.
- Data and time efficiency: 50 times fewer demonstrations and 507 times faster acquisition than the strongest task-specific fine-tuning baseline evaluated in the paper.
- Skill retention: new skills are supplied through external video context rather than written into the shared policy weights.
- Robust execution: evaluated under lighting changes, unseen objects, scene replacement, and human disturbances during execution.
..."
From the abstract:
"The ability to acquire skills rapidly and effortlessly while retaining those already mastered is essential for robots. However, current methods still rely on a cumbersome training-time loop that is costly and slow, while eroding skills already mastered.
In this paper, we introduce HOST (Human-to-robot One-Shot Skill AcquisiTion), a framework that enables a robot to acquire skills in seconds from a single human video while retaining previously mastered skills.
HOST resolves skill acquisition through a cascade of self-grounded prediction.
It first estimates the robot's progress within the demonstrated task, then translates the upcoming progression into the robot's own future observations, and finally derives actions from these predicted observations.
This cascade is trained on targets coupled to the video demonstration, obtained by mapping the robot trajectory and the video demonstration onto a shared task progress manifold, then redefining each target to align with the future progression of the video.
HOST thereby enables the robot to actively follow the demonstrated procedure and adapt it to the robot's embodiment.
HOST acquires novel skills at inference time from a single human video in an average of 29 seconds and achieves a 62% average success rate. It exceeds the zero-shot baseline by 45% while retaining previously mastered skills.
HOST even exceeds the baseline fine-tuned on 50 robot demonstrations per task while requiring 50 times fewer demonstrations and acquiring each skill 507 times faster. ..."
Project website
Robots Acquire Manipulation Skills in Seconds from a Single Human Video (preprint, open access)
No comments:
Post a Comment