本文由 AI 分析生成
建立時間： 2026-06-09 來源： https://arxiv.org/abs/2503.09018
Summary
FABCO (Feasibility-Aware Behavior Cloning from Observation) addresses a fundamental mismatch in robot imitation learning: human demonstrators may unintentionally perform motions that robots cannot physically execute. The system uses pre-trained forward and inverse dynamics models to score each demonstration segment for feasibility, provides real-time visual feedback to the demonstrator via a hand-mounted interface, and weights training samples by feasibility scores during policy learning. A user study with four participants on a pipette insertion task showed significantly higher task success rates compared to standard behavior cloning.
FABCO（可行性感知行為克隆）解決了模仿學習中人機動作差異的核心問題：人類示範者可能示範了機器人無法執行的動作。系統利用預訓練的前向與逆向動力學模型評估每個示範的可行性，透過手戴式介面提供即時視覺回饋，並在訓練時以可行性分數加權樣本。對四位受試者進行移液管插入任務的實驗，結果顯示成功率顯著優於標準行為克隆。
Prerequisites
- Imitation Learning from Observations (ILfO) — FABCO extends ILfO, which trains robot policies from state-only demonstrations without requiring action labels; understanding why action-free observations are preferable in cross-embodiment scenarios is essential.
- Behavior Cloning (BC) — the policy learning backbone; knowing how supervised regression on demonstration data works explains why infeasible samples corrupt training via gradient dilution.
- Forward and Inverse Dynamics Models — FABCO’s feasibility estimator depends on these; forward models predict next states from actions, inverse models infer actions from state transitions, so understanding their training requirements and failure modes matters.
- NASA-TLX — the workload metric used in the user study; familiarity helps interpret whether the feedback interface imposes acceptable cognitive burden.
Core Idea
The central insight is that standard behavior cloning weights all demonstrations equally, which is wrong when some demonstrations are physically infeasible for the robot. FABCO explicitly models feasibility as a continuous score derived from the robot’s own dynamics knowledge: given a demonstrated state transition, the inverse dynamics model infers the required action, and the forward dynamics model checks whether that action would actually produce the observed transition on the robot. Low agreement between the two indicates infeasibility. These scores serve two roles simultaneously — they guide the demonstrator in real time (visual feedback loop) to self-correct before data is even collected, and they reweight the training objective so that the cleaner, feasible demonstrations dominate gradient updates. This dual-use of feasibility scores — both at data collection time and at training time — is what distinguishes FABCO from simple data filtering approaches.
Results
| Task / Benchmark | This work (FABCO) | Baseline (standard BC) | Delta |
|---|---|---|---|
| Pipette insertion success rate | Significantly higher | Lower | Reported as significant improvement |
| NASA-TLX workload (demonstrator) | Acceptable | N/A (no feedback interface) | — |
Note: Exact numeric values were not available in the abstract/summary; the authors report significance but full tables require reading the complete paper.
Limitations
- Author-stated: Evaluated on a single task (pipette insertion) with only four participants; generalizability across robot morphologies and task types is unvalidated.
- Author-stated: The dynamics models must be pre-trained, requiring an initial phase of robot interaction data collection before the demonstration interface can be used.
- Unstated: The feasibility feedback loop may inadvertently constrain human demonstrations toward motions the robot already knows, potentially limiting the diversity and novelty of demonstrations that could expand the robot’s capability envelope.
- Unstated: Visual feedback of feasibility scores may not be intuitive for non-expert demonstrators; the NASA-TLX result from four participants is insufficient to generalize workload conclusions.
Reproducibility
- Code: Not mentioned in abstract; likely unavailable at time of publication (March 2025 preprint).
- Datasets: Custom pipette insertion task — no standard benchmark dataset used, limiting direct comparison.
- Compute: Not specified; dynamics model pre-training requirements depend on robot and environment complexity.
Insights
The framing of feasibility as a bridge between the human and robot’s movement spaces is conceptually important for the LfD field. Most work in ILfO focuses on the policy learning side; FABCO is unusual in also modifying the data collection interface based on robot self-knowledge. This suggests a design principle: the robot should be an active participant in shaping the demonstrations it receives, not a passive learner. The hand-mounted interface as a demonstration modality is also notable — it combines the naturalness of hand motion with the robot’s kinematic context, which is relevant to the broader trend of wearable/egocentric demonstration systems. The small user study (N=4) is a limitation but the HRI community will likely value the NASA-TLX measurement as a baseline for future interface comparisons.
Connections
- how-to-train-your-robots-demonstration-modality — directly related survey on demonstration modalities in robot learning
- egomimic-scaling-imitation-learning-egocentric-video — egocentric/hand-mounted perspective for demonstration collection
- data-scaling-laws-imitation-learning-robotic-manipulation — data quality and weighting in imitation learning at scale
- arcap-ar-robot-data-collection-system — AR-based demonstration interface with real-time feedback, complementary approach
Raw Excerpt
“Due to the differences in human and robot movement characteristics, a human expert might unintentionally demonstrate an action that the robot cannot execute. FABCO assesses the feasibility of each demonstration using the robot’s pre-trained forward and inverse dynamics models and provides visual feedback to the demonstrators, encouraging them to refine their demonstrations.”