Kamino is a GPU-based physics solver from Disney Research. The six-minute, 45-second clip shows the system training a reinforcement-learning policy on a bipedal robot that carries six nested kinematic loops, running 4,096 parallel environments on a single GPU. The description says the solver is written in Python on NVIDIA Warp and sits inside the Newton framework. The open question is whether the solver’s fidelity holds when the simulated biped is swapped for a hardware robot with its own contact noise.

A biped with six nested kinematic loops

The upload carries no narration, so the visual evidence is the only on-screen account. The description says the demo trains a reinforcement-learning policy on a biped with six nested kinematic loops. In the clip, the resulting walking policy runs in simulation. The description states that 4,096 parallel environments run on one GPU during that training. It also says Kamino supports heterogeneous worlds, which means structurally different robots can be batched onto the same GPU in one run. The publication link in the description points to kamino.disneyresearch.com. The biped, the loop count, and the environment count are all stated in the description and visible in the clip. Beyond those lines, the clip provides no spoken explanation of the solver internals.

Kinematic loops without mimic joints

The description says kinematic loops are often handled by approximating the system as a kinematic tree and adding explicit loop-closure constraints or so-called mimic joints. Kamino aims to support those couplings natively instead. The solver is integrated into the Newton framework and built on NVIDIA Warp, both of which already appear in the robot-learning toolchain. The description calls the core algorithm state-of-the-art and based on ADMM, transcribing forward dynamics as a nonlinear complementarity problem. The team’s earlier Disney Research uploads from 2026 focus on character motion, generative editing, and quadruped puppeteering. This is the first such demo from the channel that targets a rigid-body physics solver rather than a motion synthesis pipeline.

What the 6 minutes 45 seconds leave out

Nothing in the upload states whether the walking policy was tested on hardware. The clip does not state playback speed, and with no narration a viewer cannot tell whether the 4,096-environment run is shown at real time or compressed. There is no mention of retakes, cuts, or the specific GPU model used. No cost figure, energy figure, or timeline appears. The description does not say whether the biped is a specific named robot or a synthetic test model. The independent articles gathered alongside the upload cover Disney stock and Disney Plus pricing; neither mentions Kamino.

Stated in the description, absent from the clip

• Stated: The solver is implemented in Python using NVIDIA Warp and integrated into the Newton framework.
• Stated: The demo biped has six nested kinematic loops and the training run uses 4,096 parallel environments on a single GPU.
• Stated: The core algorithm is based on ADMM and transcribes forward dynamics as a nonlinear complementarity problem.
• Not stated: The specific GPU model, the hardware robot tested, or whether any run was teleoperated.
• Not stated: A publication date, a cost figure, or a timeline for availability outside the Newton framework.

From animation pipelines to robot learning

Disney Research has spent recent uploads on character animation and motion editing. Kamino points the team at the simulation layer that robot-learning groups already depend on. If the solver can batch structurally diverse robots without decomposing their loops, teams that train policies in simulation may spend less time hand-building mimic joints. Whether that holds once a policy leaves the GPU and lands on a physical biped with real contact noise is the next test to watch. The publication link is the place to check for those results.

Watch next

Most AMAZING Robots In The World! · Awesome Humanoid Robots (2014-2015)

Source: Kamino: GPU-based Simulation of Rigid Multi-Body Physics with Kinematic Loops by Disney Research , published 2026-10-12. This article was written automatically from the video above and has not been independently verified.