Verne V0.5
Collection
3 items • Updated • 1
Request access to Verne V0.5
Verne reviews each request before it gives access. Verne uses your answers to manage your license and to contact you about Verne V0.5. The Verne Privacy Notice describes how Verne processes this information.
The Verne V0.5 Model License, version 1.0, governs the use of Verne V0.5. Read the license before you request access. When you submit this form, you accept the license. For commercial use, please contact partnership@vernerobotics.com. We would love to work with you! The license does not permit use in a country or region under comprehensive US sanctions, or use by a person on a US restricted-party list.
Log in or Sign Up to review the conditions and access this model content.
Verne V0.5 RoboTwin 2.0
Verne V0.5 is an open-weight world action model for general-purpose robotics, made by Verne. This checkpoint is V0.5 fine-tuned on the RoboTwin 2.0 simulation benchmark. It succeeds in 81.0% of clean-scene episodes and 73.4% of randomized-scene episodes.
Model details
| Developed by | Verne Team |
| Model type | World action model (WAM): a video transformer and an action transformer, trained with flow matching |
| Inputs | Current camera frames, memory of past observations, robot state, and a language instruction |
| Output | A chunk of 32 targets at 30 Hz: tool poses, gripper openings, base poses, and progress |
| Instruction language | English |
| Fine-tuned from | A V0.5 pretraining checkpoint |
| Fine-tuning data | The RoboTwin 2.0 clean-scene demonstrations, 50 per task |
| License | Verne V0.5 Model License |
Results
RoboTwin 2.0 has 50 two-arm tasks. Each policy trains on 50 demonstrations per task in clean scenes. It then runs in clean scenes, and in randomized scenes with clutter, new textures, other lighting, and other table heights. Each task ran 100 episodes per scene type, except one batch of 10 randomized Scan Object episodes that did not run: 9,990 episodes in total.
| Scene type | Success | Episodes |
|---|---|---|
| Clean | 81.0% | 4,049 / 5,000 |
| Randomized | 73.4% | 3,662 / 4,990 |
Comparison with the leaderboard
| Model | Clean | Randomized |
|---|---|---|
| GO-3 Nano Preview | 88.5% | 81.5% |
| PICA-WAM | 90.7% | 74.7% |
| Verne V0.5 | 81.0% | 73.4% |
| EWAM | 82.2% | 72.1% |
| ViGAR | 82.0% | 67.0% |
| UniWAM | 75.1% | 67.6% |
| π0.5 | 70.7% | 46.0% |
| π0 | 46.4% | 16.3% |
Other results are selected from the RoboTwin 2.0 leaderboard of October 30, 2026. π0 is a single-task result. Rows are sorted by the mean of the two columns.
The lead of V0.5 over π0.5 grows from 10.3 points in clean scenes to 27.4 points in randomized scenes.
Success per task
| Task | Clean | Randomized |
|---|---|---|
adjust_bottle |
68% | 76% |
beat_block_hammer |
87% | 78% |
blocks_ranking_rgb |
97% | 93% |
blocks_ranking_size |
66% | 61% |
click_alarmclock |
100% | 98% |
click_bell |
99% | 95% |
dump_bin_bigbin |
89% | 69% |
grab_roller |
94% | 83% |
handover_block |
71% | 70% |
handover_mic |
99% | 84% |
hanging_mug |
10% | 16% |
lift_pot |
99% | 99% |
move_can_pot |
90% | 88% |
move_pillbottle_pad |
84% | 76% |
move_playingcard_away |
92% | 84% |
move_stapler_pad |
41% | 30% |
open_laptop |
72% | 59% |
open_microwave |
64% | 59% |
pick_diverse_bottles |
85% | 69% |
pick_dual_bottles |
99% | 94% |
place_a2b_left |
77% | 56% |
place_a2b_right |
66% | 42% |
place_bread_basket |
79% | 78% |
place_bread_skillet |
88% | 66% |
place_burger_fries |
95% | 85% |
place_can_basket |
82% | 83% |
place_cans_plasticbox |
84% | 80% |
place_container_plate |
99% | 94% |
place_dual_shoes |
47% | 36% |
place_empty_cup |
100% | 96% |
place_fan |
76% | 70% |
place_mouse_pad |
46% | 42% |
place_object_basket |
73% | 68% |
place_object_scale |
86% | 82% |
place_object_stand |
92% | 86% |
place_phone_stand |
92% | 73% |
place_shoe |
88% | 77% |
press_stapler |
82% | 88% |
put_bottles_dustbin |
93% | 63% |
put_object_cabinet |
41% | 24% |
rotate_qrcode |
83% | 53% |
scan_object |
65% | 61.1% |
shake_bottle |
96% | 96% |
shake_bottle_horizontally |
97% | 96% |
stack_blocks_three |
81% | 70% |
stack_blocks_two |
100% | 99% |
stack_bowls_three |
92% | 97% |
stack_bowls_two |
99% | 98% |
stamp_seal |
75% | 77% |
turn_switch |
69% | 51% |
| All 50 tasks | 81.0% | 73.4% |
Checkpoints
| Checkpoint | Description |
|---|---|
| v0.5-base | Pretrained base. Start here to fine-tune. |
| v0.5-libero | Fine-tuned on LIBERO. 99.3% average success. |
| v0.5-robotwin | Fine-tuned on RoboTwin 2.0. 81.0% clean, 73.4% randomized. |
Model
A video transformer and an action transformer share information through masked attention. During training, they learn together with video and action flow-matching objectives. At runtime, only the action path runs. The model does not generate future video.
V0.5 trains in three stages: robot-video pretraining, joint video and action pretraining, and task post-training. The pretraining data are over 1.7 million demonstrations. See the base card for details. This checkpoint completes task post-training on RoboTwin 2.0.
Uses
Intended uses
- Research on robot learning.
- Comparison with other models on the RoboTwin 2.0 benchmark.
Out of scope
- Use on a physical robot. This checkpoint is tuned for the RoboTwin 2.0 simulator.
- Any use that the license does not permit. This includes weapons, unlawful surveillance, and use in sanctioned countries and regions.
Limitations
- This checkpoint is tuned for the RoboTwin 2.0 simulator, its robot, and its cameras. Do not expect it to work on other robots or scenes without fine-tuning.
- Success in randomized scenes is lower than in clean scenes. The fine-tuning data contain clean scenes only.
- This release contains the weights and a minimal inference package. It has no training or fine-tuning code. To fine-tune V0.5 on your own robot and tasks, sign up for the Verne fine-tuning platform at app.vernerobotics.com.
Physical safety
V0.5 outputs actions that move physical machines. A machine that moves can injure people and damage property. Before you run V0.5 on a robot, assess the risks, install an emergency stop and speed and force limits, and supervise the robot while it runs. Section 6 of the license makes you responsible for these measures.
License
The Verne V0.5 Model License governs the use of these weights. In short:
- For commercial use, please contact partnership@vernerobotics.com. We would love to work with you!
- Use in a country or region under comprehensive US sanctions, and use by a person on a US restricted-party list, is not permitted.
- Weapons and unlawful surveillance are not permitted.
This summary is not the license. The LICENSE file controls. See NOTICE for third-party notices.
Citation
@misc{verne2026v05,
title = {Verne V0.5: An Open-Weight World Action Model to Accelerate Industrial Robot Deployments},
author = {{Verne Team}},
year = {2026},
url = {https://vernerobotics.com}
}
Contact
Fine-tuning platform: app.vernerobotics.com. Commercial use: partnership@vernerobotics.com. Website: vernerobotics.com.
- Downloads last month
- 5