93% Zero-Shot Grasp Rate via Skill Composition

Q: What is skill composition in humanoid robotics?

Skill composition refers to combining separately trained motion policies — such as a grasping policy and a walking policy — into a single coherent behavior at runtime, rather than training one monolithic policy for the entire task sequence. The Meng and Xiao paper uses hand-object interaction scores to arbitrate which policy controls which robot joints at each moment.

Q: What does zero-shot generalization mean for robot grasping?

In this context, zero-shot generalization means the grasping policy successfully handles objects it has never encountered during training, without any fine-tuning or additional data collection on those specific objects. The paper reports 93% success on such unseen objects.

Q: Why does training locomotion with fingers included matter?

The paper finds that even a locomotion policy, which doesn't need finger control to walk, must be trained on the full humanoid body — including fingers — for skill composition to work reliably. The authors attribute this to the joint dynamics and whole-body coordination patterns that emerge only when the complete kinematic chain is present during training.

Q: How does the finger decoupling curriculum work?

The intra-episode curriculum progressively decouples finger joints during training, drawing on developmental principles from biological motor learning. It allows the robot to start with coarser whole-hand control and gradually acquire finer individual finger control, without requiring pre-defined grasp primitives or external grasp datasets.

Q: How does this compare to end-to-end loco-manipulation approaches?

End-to-end approaches train a single policy to handle all stages of a task but typically require large datasets and can be brittle to distribution shifts. This compositional approach uses smaller, modular policies with a learned arbitration mechanism, potentially offering better sample efficiency and interpretability — though broader object and environment diversity testing would be needed to make a definitive comparison.


# Does Skill Composition Finally Crack Humanoid Loco-Manipulation?

A new paper from Fanxing Meng and Jing Xiao reports a **93% zero-shot success rate** for grasping previously unseen objects and a **96–100% success rate** for standing up while holding a grasped object — using no external datasets and no pretrained models. Published on arXiv today (2608.00208), the work directly targets one of humanoid robotics' hardest open problems: how to chain manipulation and locomotion skills together without the seams showing.

The core insight is deceptively simple but technically demanding. Rather than training a monolithic policy to grasp-then-stand-then-walk, the authors learn two separate policies — one for [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) reaching and grasping, one for getting up and walking — and compose them at runtime using hand-object interaction scores to determine which policy controls which joints at any given moment.

This is architecturally significant. Most industry approaches to [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation) either train end-to-end (expensive, brittle) or use hard-coded state machines to hand off between behaviors (inflexible). This paper proposes a middle path: modular policies with a learned arbitration signal.

---

## How the Grasping Policy Works

The reaching and grasping policy draws on harmonic analysis, using **cubic harmonics as weights** to represent the spatial relationship between the robot's hand and the target object via spatial convolution. The authors frame this as an interaction representation problem — encoding where the hand is relative to the object in a form the policy can generalize from.

Crucially, [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) of the fingers is introduced through an **intra-episode finger joint decoupling curriculum** inspired by developmental principles — roughly analogous to how biological motor development proceeds from coarse whole-limb control toward fine distal control. The robot is not handed pre-defined grasp primitives; it discovers them through structured curriculum progression.

No external datasets. No pretrained models. The 93% [zero-shot generalization](https://humanoidintel.ai/glossary/zero-shot-generalization) figure on unseen objects is the claimed payoff of this approach.

*Skeptical note:* The paper does not specify the diversity of the unseen object set, the testing environment, or whether the experiments ran in simulation, on hardware, or both. Zero-shot numbers without that context are hard to benchmark against competing approaches. Readers should treat the figures as strong but provisionally verified until replication or supplementary material clarifies conditions.

---

## The Composition Architecture

The second contribution is arguably the more strategically interesting one for the industry. The authors combine the grasping policy with a separately learned getting-up-and-walking policy by:

1. Providing each policy its own **respective observation vector** — they're not sharing a single state representation
2. Using **hand-object interaction scores** as the arbitration signal — when the hand-object score is high (active grasp), the grasping policy retains control of the relevant joints; when locomotion dominates, that policy takes over

The 96–100% success rate for standing up while holding an object is the downstream proof that this arbitration actually works. Standing up from a grasped position is notoriously difficult because it requires simultaneous re-balancing of the whole body while maintaining grasp force — a problem that trips up both rigid state machines and naive end-to-end policies.

---

## The Whole-Body Insight

Perhaps the most practically important finding the authors surface: **combining policies is only effective if each policy was trained on the whole humanoid body**, even when a given policy appears to only need a subset of the body. Their locomotion policy, for instance, doesn't obviously need finger control — but training it with fingers included turns out to matter for successful composition.

This has direct implications for how teams at [Figure AI](https://humanoidintel.ai/companies/figure-ai), [Agility Robotics](https://humanoidintel.ai/companies/agility-robotics), and others structure their policy training pipelines. If modular skill composition is the path to scalable humanoid behaviors, the "train each skill on the full body even if it seems unnecessary" constraint is a non-trivial engineering cost that needs to be baked into simulation infrastructure from the start.

---

## Why This Matters for the Industry

The field has been bifurcating. One camp — exemplified by large foundation model approaches — is betting on massive [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) with generalist policies trained on broad datasets. The other is returning to modular, compositional architectures that are more interpretable and potentially more sample-efficient.

This paper is a meaningful data point for the second camp. The claim that you can achieve high success rates on a complex sequential behavior (grasp → stand → walk) using independently trained, composable policies — without a central orchestrator and without any external training data — challenges the assumption that generalist end-to-end training is the only viable path to robust [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation).

The developmental curriculum approach also deserves attention from hardware teams. If the finger decoupling schedule is robust, it could meaningfully reduce the teleoperation data burden that companies like [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) and [Skild AI](https://humanoidintel.ai/companies/skild-ai) currently shoulder to train manipulation behaviors.

---

## Key Takeaways

- **93% zero-shot grasp success** on unseen objects, trained without external datasets or pretrained models, using cubic harmonics to represent hand-object spatial relationships
- **96–100% success** standing up while holding a grasped object, using a policy composition framework arbitrated by hand-object interaction scores
- **Whole-body training is a prerequisite for composition** — each modular policy must be trained on the full robot body for skill arbitration to work reliably
- The **intra-episode finger joint decoupling curriculum** offers a potential path to reducing reliance on large demonstration datasets for dexterous tasks
- Results need context on object diversity, environmental conditions, and sim-vs-hardware breakdown before industry teams can benchmark against current SOTA

---

## Frequently Asked Questions

**What is skill composition in humanoid robotics?**
Skill composition refers to combining separately trained motion policies — such as a grasping policy and a walking policy — into a single coherent behavior at runtime, rather than training one monolithic policy for the entire task sequence. The Meng and Xiao paper uses hand-object interaction scores to arbitrate which policy controls which robot joints at each moment.

**What does zero-shot generalization mean for robot grasping?**
In this context, zero-shot generalization means the grasping policy successfully handles objects it has never encountered during training, without any fine-tuning or additional data collection on those specific objects. The paper reports 93% success on such unseen objects.

**Why does training locomotion with fingers included matter?**
The paper finds that even a locomotion policy, which doesn't need finger control to walk, must be trained on the full humanoid body — including fingers — for skill composition to work reliably. The authors attribute this to the joint dynamics and whole-body coordination patterns that emerge only when the complete kinematic chain is present during training.

**How does the finger decoupling curriculum work?**
The intra-episode curriculum progressively decouples finger joints during training, drawing on developmental principles from biological motor learning. It allows the robot to start with coarser whole-hand control and gradually acquire finer individual finger control, without requiring pre-defined grasp primitives or external grasp datasets.

**How does this compare to end-to-end loco-manipulation approaches?**
End-to-end approaches train a single policy to handle all stages of a task but typically require large datasets and can be brittle to distribution shifts. This compositional approach uses smaller, modular policies with a learned arbitration mechanism, potentially offering better sample efficiency and interpretability — though broader object and environment diversity testing would be needed to make a definitive comparison.