# Reflex (Humanoid Control)

> Source: https://aiwiki.ai/wiki/reflex_humanoid_control
> Updated: 2026-11-08
> Categories: AI Research, Reinforcement Learning, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Reflex (Humanoid Control)." aiwiki.ai, 8 Nov 2026. https://aiwiki.ai/wiki/reflex_humanoid_control
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Reflex** is a research framework for visually reactive whole-body control of a [Unitree G1](https://aiwiki.ai/wiki/unitree_g1) humanoid. It learns to receive human-thrown boxes on its forearms and torso, absorb their momentum, and keep holding them while balancing. At deployment, an onboard RGB-D camera supplies the box observations; the controller also receives the robot state needed for stabilization.[1][2]

Taoyang Jia, Weikai Huang, and colleagues introduced the framework in *Reflex: Visually Grounded Reactive Humanoid Control*, an October 8, 2026 arXiv preprint. The project brings together researchers affiliated with the University of Washington, Stanford University, the University of Southern California, the National University of Singapore, the University of North Carolina at Chapel Hill, and the Allen Institute for AI.[1][2]

## Task and control interface

The research uses a 29-degree-of-freedom G1. A torso-mounted camera supplies 192 x 144 RGB-D observations, and the policy produces 29 joint-position commands at 50 Hz. Simulation runs through [NVIDIA Isaac Lab](https://aiwiki.ai/wiki/isaac_lab), with physics stepping at 200 Hz. No external camera measurement or ground-truth box pose is supplied to the deployed visual controller.[1]

The paper distinguishes three outcomes. Interception means slowing the incoming box near the torso. Catch success additionally requires sustained support in the chest region, with the wrists on opposite sides and without support from the head or shoulders. The box must remain at low speed for 0.6 seconds. Torso excursion and falls are separate measurements, because a successful hold can still involve substantial body tilt. The box-size experiments use a different sustained-hold criterion to accommodate different geometries.[1]

## Three-stage learning framework

Reflex separates motor control, object-dynamics inference, and visual perception. An eight-dimensional learned representation of the moving box connects the stages. Visual learning then replaces the source of that representation while retaining the trained control function.[1][3]

| Stage | Training input and objective | Result |
|---|---|---|
| Whole-body teacher | Simulator box-state observations, catching rewards, human motion tracking, and an adversarial motion prior; trained with [Proximal Policy Optimization](https://aiwiki.ai/wiki/ppo) | A controller that learns interception, impact absorption, and retention |
| Object-dynamics latent | Delayed and incomplete privileged box observations; teacher-action matching followed by joint encoder/controller PPO | A compact representation that conditions the controller |
| Visual grounding | A history of onboard RGB-D observations and camera ego-motion; supervised latent prediction | A visual encoder feeding the fixed controller |

### Teacher and latent controller

The first stage uses [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) without rendering camera images. Rewards encourage moving toward the box, supporting it near the chest, and maintaining the catch. Human motion references and an adversarial motion prior guide body movement. The teacher's box observations are corrupted with noise, missed measurements, and delay during training.[1]

The second stage initializes its new latent pathway through [imitation learning](https://aiwiki.ai/wiki/imitation_learning), matching actions from the teacher. It also collects teacher labels on states visited by the latent controller, a DAgger-style data aggregation step. PPO then updates both the privileged encoder and controller. Box-observation latency is randomized from 20 to 60 milliseconds.[1]

PPO, adversarial motion priors, and data aggregation are established methods rather than inventions of Reflex. PPO alternates environment sampling with optimization of a policy objective. AMP supplies a learned motion-style reward, while DAgger addresses the difference between expert-observed states and the states an imitating policy actually encounters.[6][8][9]

### Visual encoder

The visual encoder uses 25 frames of history. A convolutional localizer predicts a box heatmap, from which a geometric layer extracts image position, depth, confidence, and a camera-frame surface point. A recurrent model combines those features with camera ego-motion and predicts the dynamics latent. An auxiliary output predicts box position and velocity during training; the latent alone conditions the box-dependent part of control.[1]

During this stage, the controller stays fixed. The localizer is trained first, then frozen while the temporal model learns from latent targets and additional rollouts. The paper reports about 1.6 million rendered steps for visual training. This separation keeps image rendering out of the PPO loop.[1]

The fixed-controller, learned-latent approach follows an established privileged-learning direction exemplified by Rapid Motor Adaptation (RMA). RMA originally combined a base locomotion policy with a learned adaptation module for quadrupeds. Reflex applies a related separation to the perception of a fast-moving object during humanoid catching.[1][7]

## Reported results

The results below are from the authors' experiments. Simulation, real-robot catching, and an extended carrying task use different protocols and should be read separately.[1]

### Simulation

Across six trained controllers and 3,840 held-out throws, the visual system caught 3,290 boxes, or 85.7%. Giving the same latent-control architecture privileged box observations produced 87.9% success. This comparison measures the loss when a learned visual representation replaces privileged observations.[1]

| Comparison | Reported finding |
|---|---|
| Six-controller visual versus privileged input | 85.7% versus 87.9% catch success |
| Removing visual history, centered throws | 85.2% to 49.6% in the five-policy visual ablation |
| Removing visual history, 0.6 m lateral offset | 66.6% to 26.8% in the same ablation |
| Main visual checkpoint, 5 m throws | 43.0%, versus 79.7% with privileged latent input |
| Extending visual-training throws to 2.5-7 m | Simulated success at 5 m increased to 75.8%; this extension was not tested on hardware |

The visual controller did outperform the earlier state-policy reference at some lateral offsets. For one checkpoint, it reached 66.6% at 0.6 m offset, compared with 54.1% for the state-policy reference. The paper attributes this difference to additional latent-controller PPO training in stage two. It does not show that camera observations are generally more informative than perfect simulator state.[1]

The condition benchmark uses one checkpoint per controller, with 640 throws per condition for the principal comparisons. Those individual checkpoint results are distinct from the six-controller averages. The authors caution that small differences in single-checkpoint results are unresolved by those comparisons.[1]

### Real-robot trials

The real G1 trials used human throws and 20 robot trials per condition. For distance tests, the regular box measured 31 x 26 x 21 cm. The human reference pooled three people, each taking 20 throws per distance.[1]

| Throw distance | G1 successful catches | G1 catch rate | Human reference rate |
|---|---:|---:|---:|
| 2.5 m | 13 of 20 | 65% | 85.0% |
| 5 m | 8 of 20 | 40% | 68.3% |
| 7 m | 3 of 20 | 15% | 28.3% |

With only 20 robot trials per condition, the reported 95% Wilson intervals can extend roughly 20 percentage points in either direction. These are controlled research trials, not estimates of reliability in arbitrary workplaces. The project also demonstrates different box sizes, payloads, and off-center throws.[1][2]

### Carrying and handover

An additional demonstration transfers control after a settled catch to the pretrained [SONIC](https://aiwiki.ai/wiki/sonic_humanoid_control) motion controller. It carries the box to a person, hands it over, and walks back. Across 20 trials, cumulative success was 65% after catching, 35% after carrying, 30% after handover, and 25% after return. These cumulative rates make the complete sequence more demanding than the catch alone.[1]

## Software release and reproduction

The official `ToberJ/reflex` repository contains simulation training code, evaluation scripts, and released checkpoints. It builds on BeyondMimic's motion-tracking environment and runs in Isaac Lab. The onboard real-robot perception and deployment stack is excluded.[3][4]

The repository documents Linux with an NVIDIA RTX GPU. Camera rendering in stage three and visual evaluation requires RT cores, so the documented visual workflow cannot run on A100 or H100 GPUs. Its stated example memory requirement is about 14 GB for 320 visual-evaluation environments and about 6 GB for the other evaluations and training. The reported setup uses Isaac Sim 5.1.0, Isaac Lab 2.3.1, and PyTorch 2.7.0.[3]

Reproduction has limits. The released teacher resulted from several warm-started training runs whose earlier run directories no longer exist. Running its final configuration from scratch is therefore not expected to reconstruct that teacher. The supplied checkpoints let later stages and evaluations start from the released state. Camera rendering and training are also nondeterministic, so reruns can differ by several percentage points.[4]

The project code is MIT-licensed, while dependencies and motion datasets retain their own terms. The release does not redistribute the BONES-SEED reference motion or AMASS motion data; its scripts reconstruct required files from separately obtained sources.[5]

## Limitations

Reflex studies one catching task rather than general humanoid manipulation. Its principal visual-training distribution uses throws from 3.0-3.5 m. Longer throws expose a substantial perception gap, even when the localizer detects the box. The controller is also effectively insensitive to box size, and some boxes are intercepted but dropped instead of retained.[1]

The authors' hardware latency measurements came from a detector-based pipeline, not a direct timing characterization of every visual-latent component. Motion blur, auto-exposure dynamics, and structured depth artifacts are absent from the visual training model. Catch retention, torso stability, and safe deployment therefore require separate assessment.[1]

The project website's interactive viewer replays precomputed successful simulation rollouts. Its selections illustrate behaviors under different conditions; they are not new live trials and do not measure the system's success rate.[2]

## References

1. Taoyang Jia, Weikai Huang, Linxin Song, Jared Darlington, Jieyu Zhao, Yue Wang, Jiafei Duan, Zhongzheng Ren, and Ranjay Krishna. [Reflex: Visually Grounded Reactive Humanoid Control](https://arxiv.org/html/2610.12772v1). arXiv:2610.12772v1, October 8, 2026. Methods, simulation tables, real-robot evaluation, and limitations.
2. Reflex research team. [Reflex: Visually Grounded Reactive Humanoid Control, official project page](https://toberj.github.io/reflex/). Demonstrations, author affiliations, and interactive-viewer description. Accessed November 8, 2026.
3. Reflex research team. [ToberJ/reflex, official repository README](https://github.com/ToberJ/reflex). Simulation release, requirements, checkpoints, and evaluation instructions. Accessed November 8, 2026.
4. Reflex research team. [Reproducing the paper from this release](https://github.com/ToberJ/reflex/blob/main/docs/REPRODUCE.md). Checkpoint provenance, teacher-training limitations, and expected rerun variation. Accessed November 8, 2026.
5. Reflex research team. [Third-party notices](https://github.com/ToberJ/reflex/blob/main/THIRD_PARTY_NOTICES.md). Code, dependencies, robot descriptions, and motion-data licensing. Accessed November 8, 2026.
6. Stephane Ross, Geoffrey Gordon, and Drew Bagnell. [A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning](https://proceedings.mlr.press/v15/ross11a.html). AISTATS, PMLR 15:627-635, 2011.
7. Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. [RMA: Rapid Motor Adaptation for Legged Robots](https://arxiv.org/abs/2107.04034). Robotics: Science and Systems, 2021.
8. Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. [AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control](https://xbpeng.github.io/projects/AMP/). ACM Transactions on Graphics, SIGGRAPH 2021.
9. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. [Proximal Policy Optimization Algorithms](https://arxiv.org/abs/1707.06347). arXiv:1707.06347, 2017.