Riemann Dynamics Launches Riemann-1.0 World Action Model, Topping Robot Housework Benchmark with Human Video Training

Riemann Dynamics (黎曼动力), the embodied AI subsidiary of Chinese tech conglomerate Kunlun (昆仑万维), officially launched its first robot foundation model, Riemann-1.0, at WAIC 2026. The model immediately topped the RoboCasa-365 benchmark — widely considered the hardest robot housework challenge — with a 62.6% success rate, outperforming the previous SOTA by 8.4 percentage points. Most models on the leaderboard still struggle below 50%.

Riemann-1.0 is positioned as a World Action Model for general-purpose robot manipulation, fusing the VLA (Vision-Language-Action) and world model approaches into a single architecture. It inherits VLA's speed in outputting actions while retaining the world model's ability to simulate physical evolution — a hybrid direction that NVIDIA's Seattle Robotics Lab identified as the emerging mainstream for next-generation robot foundation models.

The model's defining feature is its training data strategy. Riemann-1.0 was trained on 232,000 hours of data covering 41 robot embodiments and thousands of interaction types. The core ingredient is over 200,000 hours of human first-person video capturing daily activities — cooking, folding laundry, tidying tables. Since these videos contain no robot action labels, Riemann Dynamics developed an automated pipeline using VLMs to segment actions, filter low-quality clips, reconstruct 3D hand poses, and convert human demonstrations into machine-readable training material.

Training proceeds in three stages controlled by action loss weights: stage one (weight 0.1) extracts pseudo-action signals from human video via a Latent Action Model; stage two (weight 0.5) introduces UMI and real robot data for calibration; stage three (weight 0.9) focuses on pure robot data for deployment-ready skill refinement.

Ablation studies confirmed the value of human video: on RoboCasa-365, removing human video drops performance from 62.6% to 48.2% — a 14.4-point gain. On the EgoVLA benchmark for dexterous humanoid robots, adding human video pre-training boosted long-horizon task success from 42.96% to 71.11%. The team attributes this to richer interaction semantics and task priors from human data, which improve generalization to unseen scenes far more than simply adding more robot trajectories.

In real-world tests across four household scenarios — block stacking, fabric folding, table cleaning, and kitchen organization — Riemann-1.0 achieved 85.00% average success rate and 94.43% process completion, ranking first against open-source baselines trained on identical data for identical steps, leading by 15 points. All four tasks exceeded 80% success.

Riemann Dynamics was established as Kunlun's dedicated embodied AI subsidiary. Kunlun, previously focused on AI-powered content verticals including gaming, music, and short-form video, is now making a strategic bet on physical-world AI through Riemann-1.0 — validating a training paradigm increasingly adopted across the industry: pre-train on massive open-world video, then align with minimal robot data.