# When Does Legacy Robot Data Actually Help Newer Hardware?

Legacy demonstration data collected on old robot hardware isn't just marginally useful or marginally harmful — it is actively worthless until a precise competence threshold is crossed, after which co-training performance can jump from 23.3% to 86.7% success on the same task. That's the core finding from a new paper published today on arXiv by Tao Wang, Hudson Hou, Yingdong Hu, and colleagues, and it has direct implications for every humanoid team sitting on archives of demonstration data from prior hardware generations.

The researchers studied a wheeled humanoid platform across two hardware generations, swapping both the camera and the gripper between generations while keeping the overall morphology fixed. Their central finding: co-training with legacy data follows a **three-phase, grokking-like transition** rather than a smooth, monotonic benefit curve. Below a task-specific transfer threshold, legacy data provides zero measurable lift. Cross that threshold, and the benefit is sharp. Push past saturation, and returns diminish again.

This matters immediately for teams at every major humanoid program that has shipped more than one hardware revision — which, as of mid-2026, includes nearly all of them.

---

## The Three Phases in Detail

The paper's empirical results are unusually clean for real-robot work. On a **flower insertion** task, success rate sat at 10.0% with legacy co-training when the new-hardware model had low competence — identical to the 10.0% baseline without legacy data, meaning no benefit whatsoever. Once the new-configuration policy crossed the transfer threshold, co-training rocketed performance to **86.7%**, a gain of more than 63 percentage points. On a **pen insertion** task already at high competence (85.0%), adding legacy data pushed the figure to **93.3%** — meaningful, but clearly in the diminishing-returns regime.

The authors characterize the mechanism theoretically through **gradient alignment**: when the new-hardware policy is too weak, gradients from legacy demonstrations point in conflicting directions and provide no useful training signal. Once the base policy reaches sufficient task competence, gradient directions align and legacy data acts as a strong regularizer and augmentation source. The paper also frames residual policy uncertainty as a key variable governing when the threshold is crossed.

This gradient alignment framing is significant. It connects the empirical three-phase observation to a mechanistic explanation that practitioners can interrogate — rather than treating data reuse as an empirical black box.

---

## Why Cross-Configuration Transfer Is Uniquely Hard for Humanoids

The [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) problem for humanoids is compounded by the rate of hardware iteration. Unlike industrial arms that run unchanged for a decade, humanoid programs routinely swap end-effectors, sensors, and actuator configurations between generations — sometimes within a single product cycle. Each swap potentially orphans the demonstration libraries built for the prior config.

The conventional assumption in the field has been that more data is almost always better, with cross-configuration data providing at least some marginal benefit as a diversity source. This paper directly challenges that assumption. At low competence, legacy data can be actively neutral at best — and the authors' framework suggests it could be harmful by corrupting gradient signals, though the paper's empirical results show the floor at zero gain rather than negative transfer.

The research also validates the three-phase pattern on a **mobile dual-arm watering task**, a setting that requires [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation) — coordinating whole-body movement with precise [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation). That the pattern holds in a more complex, mobile-base scenario strengthens the generalizability claim.

---

## The Practical Decision Rule

The paper derives what it calls a **phase-aware rule**: a prescriptive guide for when teams should prioritize collecting new-hardware demonstrations versus when they can safely mine legacy archives. The logic follows directly from the three-phase model — if your new-hardware policy hasn't cleared the transfer threshold on a given task, burning teleoperation hours on additional legacy data reprocessing is waste. Only once threshold competence is established does co-training pay off.

For humanoid programs managing large demonstration datasets, this has resource allocation implications. Data collection via teleoperation remains one of the most expensive line items in [Physical AI](https://humanoidintel.ai/glossary/physical-ai) development — operator time, session logistics, and annotation overhead add up fast. A principled criterion for when legacy data becomes useful is operationally valuable, not just academically interesting.

The paper's wheeled humanoid platform also matters for scope: this isn't a tabletop arm study. The platform retains humanoid morphology, and the mobile dual-arm watering validation involves coordinated arm-plus-base behavior that approximates the complexity of full bipedal humanoid manipulation deployments.

---

## Skeptical Read

A few caveats are worth flagging. The paper studies a scenario where overall morphology is held fixed and only camera and gripper are swapped — a relatively controlled form of cross-configuration transfer. The harder case, which this paper does not address, is transfer across more significant morphological changes: different kinematic chains, different [degrees of freedom](https://humanoidintel.ai/glossary/degrees-of-freedom) counts, or different actuator modalities (e.g., from hydraulic to [tendon-driven](https://humanoidintel.ai/glossary/tendon-driven)). The gradient alignment account may hold in those settings, but the transfer threshold location will almost certainly shift.

The task set is also limited to relatively precise insertion and watering tasks. How the three-phase pattern manifests on tasks with broader action distributions — pick-and-place in clutter, bimanual assembly — is an open question. The theoretical framework predicts the pattern should generalize, but real-robot validation across a wider task suite would significantly strengthen the claim.

Finally, the paper is silent on how to *measure* whether you've crossed the transfer threshold short of running the full co-training experiment. A practical diagnostic — a proxy metric teams could evaluate without burning full training runs — would be the natural next contribution.

---

## Industry Trajectory

The broader implication is that humanoid programs need **curriculum-aware data strategies**, not just large data pipelines. The framing of [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) across hardware configs — the implicit hope behind many large-scale data collection efforts — may be more fragile than assumed. Emergent transfer, which arrives non-linearly, requires practitioners to think about data reuse as a function of current policy state, not just data volume or coverage.

For teams building foundation models for manipulation — the territory occupied by [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) and [Skild AI](https://humanoidintel.ai/companies/skild-ai) — this research adds texture to the question of how cross-robot, cross-configuration training corpora should be constructed. If the transfer threshold phenomenon holds at scale, naively mixing legacy and new-hardware data in a single training run without conditioning on per-task competence levels could suppress performance on exactly the tasks where the legacy data would otherwise help most.

---

## Key Takeaways

- Legacy demonstration data follows a **three-phase transfer pattern**: zero benefit at low competence, sharp gain after a transfer threshold, diminishing returns near saturation.
- Empirical results on real hardware show success jumping from **23.3% to 86.7%** on flower insertion after the threshold is crossed.
- The mechanism is explained via **gradient alignment**: conflicting gradients from legacy data provide no useful signal until the new-hardware policy achieves minimum task competence.
- The pattern holds on both precision manipulation tasks and a more complex **mobile dual-arm watering** task.
- The paper derives a **phase-aware decision rule** for when to collect new-hardware data versus reuse legacy demonstrations.
- The finding directly challenges the assumption that more cross-configuration data is monotonically beneficial.
- Significant open questions remain: morphologically diverse transfer, broader task distributions, and practical threshold diagnostics.

---

## Frequently Asked Questions

**What is the transfer threshold in cross-configuration robot learning?**
The transfer threshold is a task-specific competence level that a new-hardware robot policy must reach before legacy demonstration data from prior hardware configurations provides any training benefit. Below this threshold, co-training with legacy data yields no measurable performance improvement. The concept was formalized in a July 2026 arXiv paper studying a wheeled humanoid platform across two hardware generations.

**Why doesn't legacy robot data always help when co-training?**
When a new-hardware policy has low task competence, gradients from legacy demonstrations point in misaligned directions relative to gradients from new-hardware data. These conflicting gradient signals cancel out, leaving the policy no better off than training on new-hardware data alone. This gradient alignment breakdown is the mechanistic explanation proposed in the paper.

**What is grokking in robot learning?**
Grokking refers to a phenomenon, originally observed in language model training, where a model shows no improvement for an extended period and then suddenly generalizes sharply. In the context of this paper, legacy data exhibits a grokking-like transition: it remains ineffective until a competence threshold is crossed, at which point co-training performance rises sharply.

**How does camera or gripper swapping affect imitation learning data reuse?**
Changing a camera changes the visual observation distribution; changing a gripper changes the action morphology and [end-effector](https://humanoidintel.ai/glossary/end-effector) dynamics. Both create a distribution mismatch between legacy and new-hardware demonstrations. This paper shows that mismatch is not uniformly damaging — it becomes damaging only when the new policy lacks sufficient competence to align gradient signals across the two data distributions.

**Does this research apply to full bipedal humanoids, or only wheeled platforms?**
The experiments were conducted on a wheeled humanoid with dual arms, with morphology held fixed across hardware generations. The paper's authors validated the three-phase pattern on a mobile manipulation task, suggesting robustness beyond static benchtop settings. Whether the same threshold dynamics hold for full bipedal humanoids with more complex kinematic changes remains an open research question.