# Does AdaPT Give Humanoid Robots Professional Tennis Form?

A new hierarchical framework called AdaPT — Adaptive Motion Planning and Tracking — extracts professional tennis serving and rally styles directly from broadcast video and deploys them on two physical humanoid platforms: the [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) G1 and the full-size [Dobot](https://humanoidintel.ai/companies/dobot) Atom (1.7 m tall), the latter without any motion capture infrastructure. Published on arXiv on August 21, 2026 by a team of 18 researchers, the work tackles one of the harder open problems in humanoid sports robotics: maintaining professional motion aesthetics — not just task success — while surviving the [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) gap.

The core result is a working in-the-wild tennis serve on a full-size humanoid robot, achieved without motion capture and without hand-engineered reference trajectories. The system learns from broadcast video, a data source that is essentially unlimited and requires no lab instrumentation. That matters more than the tennis itself: it demonstrates a viable pipeline for acquiring stylistically rich human motor skills at scale and deploying them on real hardware.

---

## What AdaPT Actually Does

The framework is hierarchical. A **planner** generates stylistic kinematic motions learned from broadcast tennis footage. A **tracker** executes those motions on real hardware. The design philosophy is deliberate: keep planning and execution as decoupled as possible so that stylistic quality — the thing humans perceive as "professional" — is not eroded by low-level control noise.

This is a meaningful architectural choice. Most prior work in humanoid [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) conflates style and task performance into a single policy. When that monolithic policy encounters real-hardware dynamics it wasn't trained on, both degrade together and it's hard to diagnose why. AdaPT's separation lets the team identify exactly where breakdown occurs.

The paper identifies three compounding failure modes in the sim-to-real pipeline:

1. **Tracking degradation** — real actuators can't perfectly follow simulated kinematic targets, particularly for high-speed, high-amplitude motions like a tennis serve.
2. **Autoregressive error accumulation** — the planner, which generates motions sequentially, partially overlooks tracking failures because it doesn't receive clean feedback about execution quality.
3. **Noisy perception** — real-world ball tracking introduces uncertainty that simulation does not faithfully replicate.

To address all three, the team introduces an adaptation mechanism: policies learn to track **randomized execution speeds** during training, building robustness to the range of tracking fidelity seen on real hardware. Simultaneously, a **motion-speed adapter** conditions the planner, allowing it to adjust its kinematic targets based on how well the tracker is actually performing. This closes the loop between planning quality and execution reality without requiring the planner to manage low-level control details directly.

---

## Why Broadcast Video as the Data Source Matters

The choice to learn from broadcast video rather than motion capture is strategically significant. Motion capture requires instrumented labs, cooperative human subjects, and significant per-session cost. Broadcast tennis footage — from Wimbledon, the US Open, or ATP tour matches — represents decades of professional human performance, already filtered by competition for stylistic and biomechanical quality.

The challenge is that broadcast video provides 2D observations of inherently 3D motions, and camera angles vary. Lifting professional tennis kinematics from this source to a 3D humanoid skeleton is a non-trivial perception and inverse kinematics problem. The paper does not detail the specific pose estimation pipeline used, but the downstream results on real hardware suggest it is robust enough to serve as a viable training signal.

For the broader humanoid field, this points toward a general strategy: rather than building bespoke motion datasets for each task, mine the enormous archive of human performance video that already exists for sports, dance, and skilled trades.

---

## Hardware: G1 and the Dobot Atom

The paper validates on two distinct platforms, which strengthens the generalization claim.

The **Unitree G1** is an increasingly common research testbed — cost-accessible, with a known dynamics model and a growing body of published work, making comparisons tractable.

The **Dobot Atom** (1.7 m) is the more interesting deployment target here. It is a full-size humanoid, and the paper explicitly highlights that AdaPT policies were deployed on it **without motion capture** — in-the-wild conditions. The Dobot Atom is Dobot's entry into the humanoid space, and this appears to be a meaningful real-world validation of that platform's control capabilities.

Running the same framework on two physically distinct robots with different kinematics and actuator characteristics, without platform-specific re-training details mentioned in the abstract, suggests the adaptation mechanism generalizes across hardware — though the paper's full results section would need scrutiny to confirm the degree of cross-platform transfer.

---

## Where the Skepticism Lives

Several questions the abstract leaves open that engineers and investors should press on:

**What is the baseline for "professional style"?** The paper claims to learn professional motion styles, but quantifying style is notoriously hard. Human perceptual studies, biomechanical metrics, or comparison to labeled professional reference motions would all be appropriate, and it's unclear which the paper uses.

**How does task performance trade off against style?** The framing explicitly acknowledges the tension. If stylistic motions reduce ball placement accuracy, that's a real limitation for any competitive application — and for the broader [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) research agenda where task success is the primary metric.

**Generalization beyond serve and rally?** Tennis has a constrained motion vocabulary relative to general manipulation tasks. The extent to which AdaPT's architecture transfers to less periodic, less well-defined human motor behaviors remains to be demonstrated.

**Real-time performance?** High-speed racquet sports demand low-latency perception-to-action loops. The abstract doesn't address inference latency, which is often the practical bottleneck for dynamic sports applications.

---

## Industry Trajectory Implications

Ball sports have become a legitimate benchmark category for humanoid capabilities precisely because they stress-test the properties that actually matter for deployment: dynamic motion, precise timing, real-world perception under uncertainty, and graceful degradation when conditions deviate from training. A robot that can serve a tennis ball with professional form is demonstrating [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation) capabilities under time pressure — arguably harder than most warehouse pick-and-place tasks.

The AdaPT framework's approach — hierarchical planning/tracking separation, video-based style learning, and explicit sim-to-real adaptation mechanisms — is directly applicable to industrial scenarios where motion quality, not just task completion, affects downstream outcomes. Think of precision assembly tasks where the trajectory of approach matters as much as final placement accuracy.

The paper's release of code and videos on the project website raises the value for the community: researchers can interrogate the implementation rather than just the abstract claims.

For Dobot, the Atom's appearance in a peer-reviewed arXiv paper with real-world in-the-wild results is meaningful credentialing in a market where hardware platforms compete partly on demonstrated research capability.

---

## Key Takeaways

- **AdaPT** is a hierarchical motion planning and tracking framework that learns professional tennis styles from broadcast video, requiring no motion capture data collection.
- Deployed on two real humanoid platforms: **Unitree G1** and the **Dobot Atom** (1.7 m), with in-the-wild serving demonstrated on the Atom.
- Three sim-to-real failure modes are identified and addressed: tracking degradation, autoregressive error accumulation, and noisy perception.
- The adaptation mechanism — randomized execution speed training plus a motion-speed adapter — is the core technical contribution for bridging the sim-to-real gap.
- Broadcast video as a training data source is a scalable alternative to motion capture for acquiring high-quality human motion priors.
- Code and videos are publicly available, enabling community validation.
- Open questions remain around style quantification, task-style tradeoffs, and real-time performance at competitive speeds.

---

## Frequently Asked Questions

**What is the AdaPT framework for humanoid robots?**
AdaPT (Adaptive Motion Planning and Tracking) is a hierarchical control framework that learns professional tennis serving and rally styles from broadcast video. A planner generates stylistic kinematic motions; a tracker executes them on real hardware. An adaptation mechanism bridges the sim-to-real gap by training on randomized execution speeds and conditioning the planner on a motion-speed adapter.

**Which humanoid robots were used to test AdaPT?**
The framework was validated on the Unitree G1 and the Dobot Atom, a full-size humanoid robot measuring 1.7 meters tall. The Dobot Atom deployment was conducted in-the-wild without motion capture equipment.

**Why use broadcast video instead of motion capture for robot training?**
Broadcast sports footage provides access to decades of professional human performance without requiring instrumented labs or cooperative subjects. It is a scalable, low-cost data source for learning high-quality human motor skills — particularly relevant for dynamic sports behaviors.

**What is the sim-to-real gap in humanoid sports robotics?**
The sim-to-real gap refers to the performance drop that occurs when a policy trained in simulation is deployed on real hardware. For dynamic tasks like tennis, this gap is especially pronounced due to actuator limitations, real-world perception noise, and the high-speed, high-amplitude motions required. AdaPT specifically targets this gap with its adaptation mechanism.

**How does AdaPT handle the tension between motion style and task performance?**
AdaPT's hierarchical design — separating the stylistic planner from the task-executing tracker — is intended to preserve motion quality without sacrificing reliability. The adaptation mechanism allows the planner to adjust based on tracker performance, reducing compounding errors that would otherwise degrade both style and task success simultaneously.