BoT-Feedback
An AI motion feedback generation framework grounded in biomechanical reasoning. It employs a four-stage progressive reasoning chain and a biomechanical data parser to enable multimodal large language models to produce interpretable and physically plausible motion coaching suggestions.
Timeline (last 90 days)
BoT-Feedback proposes a four-stage reasoning framework called Biomechanics of Thought (BoT).
The research team designed a plug-and-play Biomechanical Data Parser (BDP) that converts raw video into structured biomechanical descriptors.
The research team designed an alignment strategy to match expert and student actions in the temporal dimension.
The paper conducted experiments on twelve open-source and closed-source MLLMs.
BoT-Feedback improves the average expert score from 2.07 to 2.95, a relative improvement of approximately 40%.
Grounding reasoning on biomechanical evidence improves feedback quality, interpretability, and robustness, and substantially reduces biomechanical hallucinations.
The BoT-Feedback method enables compact open-source MLLMs to approach larger proprietary systems on the task of interpretable action feedback generation, narrowing the gap between open-source and closed-source models.
BoT-Feedback is a preprint paper published on arXiv.
All Facts (8)
BoT-Feedback proposes a four-stage reasoning framework called Biomechanics of Thought (BoT).
50%UnverifiedThe research team designed a plug-and-play Biomechanical Data Parser (BDP) that converts raw video into structured biomechanical descriptors.
50%UnverifiedThe research team designed an alignment strategy to match expert and student actions in the temporal dimension.
50%UnverifiedThe paper conducted experiments on twelve open-source and closed-source MLLMs.
50%UnverifiedBoT-Feedback improves the average expert score from 2.07 to 2.95, a relative improvement of approximately 40%.
50%UnverifiedGrounding reasoning on biomechanical evidence improves feedback quality, interpretability, and robustness, and substantially reduces biomechanical hallucinations.
50%UnverifiedThe BoT-Feedback method enables compact open-source MLLMs to approach larger proprietary systems on the task of interpretable action feedback generation, narrowing the gap between open-source and closed-source models.
50%UnverifiedBoT-Feedback is a preprint paper published on arXiv.
50%