Seoul National University's Dynin-Robotics hits 78.4% on a real Franka arm, edging pi0.5 and NVIDIA's GR00T
A Seoul National University lab says its Dynin-Robotics model averaged 78.4% success across four real manipulation conditions on a Franka Research 3 arm, ahead of pi0.5 (76.9%) and NVIDIA's GR00T-N1.6 (73.0%) in the same test.
The preprint, accepted to NeurIPS 2026, folds instruction following, scene prediction and action generation into one masked-diffusion model.
Seoul National University's AIDAS lab, led by Professor Do Jae-young of the Department of Electrical and Computer Engineering, reports that its robot foundation model Dynin-Robotics averaged 78.4% success across four real-world manipulation conditions on a Franka Research 3 arm. In the same test the team measured 76.9% for Physical Intelligence's pi0.5, 73.6% for NVIDIA's Cosmos Policy, 73.0% for NVIDIA's GR00T-N1.6 and 71.1% for Mimic-Video, according to the paper posted on arXiv and the project page, which lists the work as accepted to NeurIPS 2026. Korean media carried the university's announcement on Oct 17; the conference takes place in December.
The margin is small, and where it comes from is the interesting part. Dynin-Robotics beat pi0.5 by 1.0, 2.0 and 4.5 percentage points on fruit pick-and-place, cube sorting and colour-ordered stacking, and trailed it by 1.5 points on free stacking. Its clearest lead is on the one task where the instruction specifies an order.
One model that predicts the scene before it moves
Most vision-language-action models bolt an action head onto a vision-language backbone, as pi0.5 and GR00T do, or condition a policy on a video generator, as Cosmos Policy and Mimic-Video do. Dynin-Robotics is built on Dynin-Omni, the lab's own omnimodal masked-diffusion model released in March, and puts language, camera frames, goal images and quantised robot actions into one token sequence. By masking different spans, the same network learns four jobs: generate the next actions, predict the next frame given an action, imagine the finished scene, and describe a trajectory in words.
At run time those jobs can be combined. The robot can first draw the goal state it is aiming for, denoise actions and the next frame together, or use its world model to score several candidate actions and keep the best. The paper reports that goal guidance plus joint action and next-frame denoising improved success on shifted instructions over action-only decoding, while either component alone did worse, so the benefit depends on how the predictions are composed.
The numbers behind the claim
The model was continually pretrained on about 1.33 million trajectories from 48 Open X-Embodiment datasets, then adapted separately to each benchmark. It scores 98.1% on LIBERO, level with X-VLA and 0.5 points below ABot-M0, and 73.0% zero-shot on the perturbed LIBERO-Plus suite, below ABot-M0's 81.6%; camera and robot perturbations remain weak spots at 59.8% and 48.2%. An accelerated decoder built on the dInfer stack raised model-side action throughput 29.15 times to 268.8 action units per second on one NVIDIA B200, against 7.35 for pi0.5 under the same protocol. The team says weights and code will be released.
- Dynin-Robotics (SNU)
- 97.5
- 85.5
- 62.0
- 68.5
- 78.4
- pi0.5 (Physical Intelligence)
- 96.5
- 83.5
- 63.5
- 64.0
- 76.9
- Cosmos Policy (NVIDIA)
- 98.0
- 87.0
- 67.5
- 42.0
- 73.6
- GR00T-N1.6 (NVIDIA)
- 91.0
- 78.0
- 59.5
- 63.5
- 73.0
- Mimic-Video
- 96.5
- 84.5
- 64.0
- 39.5
- 71.1
Source: arXiv 2609.13053, Table 5. All models adapted and run by the SNU team on one workspace; rollout count per condition not stated.
As of Oct 17, 2026
MEKAWIRE analysis
Korean academia now builds robot foundation models that match U.S. company models on public tests; the constraint is no longer architecture but data and deployment.
MEKAWIRE's database holds two Korean robot foundation model announcements this year. Tomorrow Robotics reported 99.55% on LIBERO for Habilis Brain 0 on Oct 7, a company figure carried by Korean outlets. Dynin-Robotics reports 98.1% in a paper with full tables. Both sit in the same saturated band as pi0.5 (96.9%) and GR00T-N1.6 (97.0%), which means LIBERO no longer separates models. Real-robot results and perturbation tests do, and there the Seoul model is competitive rather than dominant.
- Habilis Brain 0
- Tomorrow Robotics (KR)
- 99.55
- Company figure as reported by Korean outlets, Oct 7
- Dynin-Robotics
- Seoul National University (KR)
- 98.1
- Paper, Table 2
- X-VLA
- Academic (CN)
- 98.1
- As listed in the SNU paper
- GR00T-N1.6
- NVIDIA (US)
- 97.0
- As listed in the SNU paper
- pi0.5
- Physical Intelligence (US)
- 96.9
- As listed in the SNU paper
MEKAWIRE compilation from the SNU paper and its database. LIBERO scores near 100% no longer separate models well.
As of Oct 17, 2026
The strongest counter-argument is scale. pi0.5 and GR00T are trained on proprietary fleets far larger than 1.33 million public trajectories, and they ship to customers; a university model tuned per benchmark on one Franka setup has not yet faced a factory. The paper does not state the number of real-world rollouts per condition, so the 1.5-point lead cannot be tested for significance.
The bull case is that the unified design gives Korean integrators such as Rainbow Robotics and Doosan Robotics a domestic, open model that can plan with imagined goal images, useful for multi-step kitting. The bear case is that without a data engine the model stays a benchmark entry. Signals to watch: release of weights and code, any Korean manufacturer naming it in a pilot, and the NeurIPS presentation in December.
- Up to 29.15x faster model-side action decoding with dInfer on one NVIDIA B200
- Accepted to NeurIPS 2026 (December); arXiv 2609.13053
- 98.1% average; LIBERO-Plus zero-shot 73.0%
- Weights and code announced as coming soon
- 78.4% average over four conditions on Franka Research 3 (pi0.5 76.9%, Cosmos Policy 73.6%, GR00T-N1.6 73.0%, Mimic-Video 71.1%)
- Seoul National University, AIDAS Lab (Prof. Do Jae-young)
- About 1.33 million trajectories from 48 Open X-Embodiment datasets
Why it matters
A Korean university has produced a robot foundation model that holds its own against Physical Intelligence and NVIDIA on a shared real-robot test, and does it with a distinct architecture rather than a fine-tuned copy. For Korean policy makers funding K-humanoid and physical-AI programmes, it is evidence that domestic model capability exists at the research level.
It also shows where the frontier has moved. Benchmarks such as LIBERO are saturated; perturbation robustness, real-robot task mix and inference speed are now the axes on which models differ.
Rival analysis
Against pi0.5, Dynin-Robotics wins three of four real conditions by 1 to 4.5 points and loses one. Against Cosmos Policy, it loses three conditions but wins colour-ordered stacking by 26.5 points, which drives its higher average. Against ABot-M0 it trails by 8.6 points on LIBERO-Plus. Domestically, Tomorrow Robotics claims a higher LIBERO score but has published no paper MEKAWIRE could check.
Valuation context
There is no company or valuation attached; the AIDAS lab is academic. The relevant comparison is capital: Physical Intelligence and NVIDIA fund data collection at a scale no Korean lab matches, and the model's commercial path depends on whether a Korean robot maker or platform company licenses or spins it out.
Supply-chain implications
Training and profiling ran on NVIDIA hardware (B200 for the throughput tests), and the real robot is a German Franka Research 3. The model therefore depends on the same U.S. compute stack as its rivals; the Korean contribution is the model and training recipe.
Signals to watch
Weights and code release (announced as soon); the NeurIPS 2026 presentation in December; any Korean manufacturer or integrator naming Dynin-Robotics in a pilot; a follow-up with more embodiments or force sensing, which the architecture reserves but this run did not use.
Analyst view
We think the result is credible and modest: a competitive model with a clear edge on language-ordered tasks and a large speed-up from parallel decoding. Confidence medium, because rollout counts are not stated and every benchmark used separate adaptation.
Questions you should be asking
How many rollouts per condition underlie Table 5? Will the released weights include the real-robot adapter? Is any Korean robot maker evaluating the model? How does it perform on a mobile or dual-arm platform?