When Does Legacy Data Start to Help? Emergent Transfer in Cross-Configuration Robot Learning

Figure 1: The three-phase pattern of cross-configuration co-training. We compare policies trained on legacy data D₁, new-hardware data D₂, and their mixture. We denote the standalone success rate on the new configuration by SR₂ and define the co-training gain as ΔSR = SR_co-train - SR₂. As SR₂ increases, legacy data is ineffective below the task-dependent transfer threshold τ(T), produces its largest gains just above the threshold, and yields diminishing returns at high baselines.

Abstract

Robotic hardware evolves over time, but demonstration data is often tied to a specific sensor and actuator configuration. This raises a practical and underexplored question: when does legacy data begin to benefit an upgraded robot? We study this question on a wheeled humanoid platform across two hardware generations, where both the camera and gripper are changed while the overall morphology remains fixed. Contrary to the common assumption that more cross-configuration data is always helpful, we observe a grokking-like transition: legacy data remains ineffective until the upgraded configuration acquires a minimum level of task competence, after which co-training gains rise sharply before diminishing near saturation. We hypothesize that this task-dependent transition is governed by a transfer threshold and characterize the resulting three-phase pattern. Across real-robot manipulation tasks, we observe all three phases: no measurable benefit at low competence (10.0% → 10.0%), a sharp gain after crossing the threshold (23.3% → 86.7% on flower insertion), and diminishing returns at high competence (85.0% → 93.3% on pen insertion). We provide a theoretical account based on gradient alignment and residual policy uncertainty, and derive a phase-aware rule for deciding when to collect more new-hardware data and when to reuse legacy demonstrations. We further validate this three-phase pattern on a mobile dual-arm watering task, with results consistent with our predictions.

Key Results

Flower Insertion

+63.4%

Success rate gain
(23.3% → 86.7%)

Data Efficiency

5.3×

Less new data needed
(8h → 1.5h)

Water Plant

+38.3%

Held-out task gain
(40.0% → 78.3%)

Key Contributions

Emergent Three-Phase Transfer Pattern: We identify and statistically support that legacy data is ineffective at low target competence, highly beneficial at intermediate competence, and subject to diminishing returns near saturation.
Theoretical Framework: We model this transition using a task-dependent transfer threshold τ(T) and provide a theoretical account based on gradient alignment, residual policy uncertainty, and domain adaptation.
Phase-Aware Data Collection: We formulate a practical rule that collects new-hardware demonstrations until the transfer threshold is crossed, then introduces legacy data for co-training, reducing collection time from 8 hours to 1.5 hours on a held-out task.

Experimental Results

Real-Robot Manipulation Tasks

We begin with a suite of real-robot manipulation tasks on a wheeled humanoid platform, evaluated across two hardware generations that differ in both camera and gripper while keeping the overall morphology fixed. These tasks span a range of standalone success rates, from near-random to near-saturation, which lets us observe how the benefit of legacy data changes as the upgraded configuration becomes more competent.

Real-robot experimental setup. We evaluate our approach on fine-grained manipulation tasks including flower insertion and pen insertion, using a wheeled humanoid platform across two hardware generations with different cameras and grippers.

The Emergent Three-Phase Pattern

Aggregating results across these tasks reveals a consistent pattern. Plotting the co-training gain ΔSR against the standalone success rate SR₂ exposes three distinct phases: below the task-dependent transfer threshold τ(T), legacy data provides no measurable benefit; just above the threshold, it produces its largest gains; and near saturation, the returns diminish. This curve is the empirical backbone of our analysis and motivates the phase-aware data collection rule.

Phase transition curve showing transfer threshold

Phase transition curve. The co-training gain ΔSR exhibits a characteristic three-phase pattern as the standalone success rate SR₂ increases. The transfer threshold τ(T) marks the transition point where legacy data begins to provide substantial benefits.

Validation: Mobile Dual-Arm Watering

Finally, we put the three-phase pattern to the test on a held-out task the curve was never fit to: a mobile dual-arm watering task. Guided by the phase-aware rule, we collect new-hardware demonstrations only until the transfer threshold is crossed, then bring in legacy data for co-training. The result matches our prediction, with success rising from 40.0% to 78.3% while cutting new-hardware data collection from 8 hours to just 1.5 hours, offering strong validation that the pattern generalizes beyond the tasks used to derive it.

Validation on a held-out watering task. The phase-aware strategy raises success from 40.0% to 78.3% while reducing new-hardware data collection from 8 hours to just 1.5 hours, consistent with the three-phase pattern derived from the manipulation tasks above.

Citation

@inproceedings{wang2027legacy,
  title={When Does Legacy Data Start to Help? Emergent Transfer in Cross-Configuration Robot Learning},
  author={Wang, Tao and Hou, Hudson and Hu, Yingdong and Liu, Yufeng and Li, Qinghai and Jiang, Yingjie and Wang, Yingzhi and Ma, Cheng and Wang, Richard and Gao, Yang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2027}
}