Version scope and roles for these associations await verification.
LocoFormer is Skild AI’s generalist locomotion policy for legged and wheeled robots. Large-scale PPO on procedurally generated robots and Transformer-XL memory enable adaptation through interaction history across trials.
Locomotion research with one policy across unseen morphologies and changed dynamics. The policy uses robot-state history and outputs target joint positions in a unified joint space; it is not a video-language manipulation model. The paper reports real-robot demonstrations including Go2/Go2-W, G1 and H1, while Table1 quantitatively evaluates ten robot morphologies in simulation. Few-shot evaluation allows five seconds of prior adaptation; initial trials can fail. September 28, 2025 is the first arXiv disclosure, not a verified public weight release.
Paper available
Last checked: 2026-11-07
Source publication date: 2025-09-28
Source publication date: 2025-09-28
Paper available
Author order: Original order
Original LocoFormer generalist locomotion policy; simulation quantitative evaluation and separate real-robot demonstrations. Not S1 manipulation or a released public checkpoint.
SourceVerified on · 2026-11-07
Skild AI / LocoFormer authors · Research paper
Published: 2025-09-28 · Accessed: 2026-11-07
LocoFormer v1: exact research protocols
Skild AI / LocoFormer authors · Research paper
Published: 2025-09-28 · Accessed: 2026-11-07
LocoFormer v1: exact research protocols
Skild AI / LocoFormer authors · Research paper
Published: 2025-09-28 · Accessed: 2026-11-07
LocoFormer v1: exact research protocols
Skild AI / LocoFormer authors · Research paper
Published: 2025-09-28 · Accessed: 2026-11-07
LocoFormer v1: exact research protocols
Skild AI / LocoFormer authors · Research paper
Published: 2025-09-28 · Accessed: 2026-11-07
LocoFormer v1: exact research protocols
Skild AI / LocoFormer authors · Research paper
Published: 2025-09-28 · Accessed: 2026-11-07
Source category does not verify the claims it contains.
No verified independent benchmark results are recorded for this version or the selected weights.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · G1 / Zero-shot ↗Normalized goal displacement [0,1]: 0.96 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. G1 / Zero-shot: 0.96 ± 0.16.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · H1 / Zero-shot ↗Normalized goal displacement [0,1]: 0.98 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. H1 / Zero-shot: 0.98 ± 0.12.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · GR1 / Zero-shot ↗Normalized goal displacement [0,1]: 0.95 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. GR1 / Zero-shot: 0.95 ± 0.21.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · TRON1 / Zero-shot ↗Normalized goal displacement [0,1]: 0.87 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. TRON1 / Zero-shot: 0.87 ± 0.31.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · BK-H / Zero-shot ↗Normalized goal displacement [0,1]: 0.98 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. BK-H / Zero-shot: 0.98 ± 0.11.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · A1 / Zero-shot ↗Normalized goal displacement [0,1]: 0.92 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. A1 / Zero-shot: 0.92 ± 0.26.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · Spot / Zero-shot ↗Normalized goal displacement [0,1]: 1 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. Spot / Zero-shot: 1.00 ± 0.06.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · AnyMal C / Zero-shot ↗Normalized goal displacement [0,1]: 0.95 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. AnyMal C / Zero-shot: 0.95 ± 0.22.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · TRON1-W / Zero-shot ↗Normalized goal displacement [0,1]: 0.99 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. TRON1-W / Zero-shot: 0.99 ± 0.07.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · Go2-W / Zero-shot ↗Normalized goal displacement [0,1]: 0.97 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. Go2-W / Zero-shot: 0.97 ± 0.16.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · 10-robot average / Zero-shot ↗Normalized goal displacement [0,1]: 0.96 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. 10-robot average / Zero-shot: 0.96 ± 0.19.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · G1 / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.98 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. G1 / 5s prior adaptation: 0.98 ± 0.11.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · H1 / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.99 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. H1 / 5s prior adaptation: 0.99 ± 0.08.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · GR1 / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.99 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. GR1 / 5s prior adaptation: 0.99 ± 0.09.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · TRON1 / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.96 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. TRON1 / 5s prior adaptation: 0.96 ± 0.18.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · BK-H / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.99 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. BK-H / 5s prior adaptation: 0.99 ± 0.09.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · A1 / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.94 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. A1 / 5s prior adaptation: 0.94 ± 0.23.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · Spot / 5s prior adaptation ↗Normalized goal displacement [0,1]: 1 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. Spot / 5s prior adaptation: 1.00 ± 0.05.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · AnyMal C / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.96 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. AnyMal C / 5s prior adaptation: 0.96 ± 0.19.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · TRON1-W / 5s prior adaptation ↗Normalized goal displacement [0,1]: 1 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. TRON1-W / 5s prior adaptation: 1.00 ± 0.04.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · Go2-W / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.98 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. Go2-W / 5s prior adaptation: 0.98 ± 0.10.
Publisher evaluation · Reported; not independently reproduced
LocoFormer unseen-robot simulation · 10-robot average / 5s prior adaptation ↗Normalized goal displacement [0,1]: 0.98 score
Author-reported Table1, not independently reproduced. Ten previously unseen robots spanning bipeds, quadrupeds and wheeled robots;1,000 simulation environments per robot with rough terrain and extensive domain randomization. Metric is displacement toward a randomly sampled goal normalized to [0,1], not percentage success, full-task autonomy or a real-hardware result. Zero-shot and five-second prior-adaptation conditions are separate. The table’s reported ± term is retained without guessing its statistical definition. No public checkpoint is identified. 10-robot average / 5s prior adaptation: 0.98 ± 0.13.
Locomotion action control
Target joint positions in a unified joint space
Locomotion research with one policy across unseen morphologies and changed dynamics. The policy uses robot-state history and outputs target joint positions in a unified joint space; it is not a video-language manipulation model. The paper reports real-robot demonstrations including Go2/Go2-W, G1 and H1, while Table1 quantitatively evaluates ten robot morphologies in simulation. Few-shot evaluation allows five seconds of prior adaptation; initial trials can fail. September 28, 2025 is the first arXiv disclosure, not a verified public weight release. No official public training-code, weight, hosted-API or model-license access has been established; the paper’s CC BY4.0 license does not grant rights to model code or weights. Table1 reports normalized displacement toward a goal on [0,1], not percentage success, across 1,000 simulated environments per robot. Results are author-reported, not independently reproduced. Training is resource-intensive and relies on a hand-crafted procedural task space. Sudden morphology changes can cause falls or failed adaptation; memory across attempts is not uninterrupted fault tolerance or deployment safety. Manipulation and general whole-body task capabilities are not established. [Paper](https://arxiv.org/abs/2509.23745) · [Project](https://generalist-locomotion.github.io/)
Before you use this version: No official public training-code, weight, hosted-API or model-license access has been established; the paper’s CC BY4.0 license does not grant rights to model code or weights. Table1 reports normalized displacement toward a goal on [0,1], not percentage success, across 1,000 simulated environments per robot. Results are author-reported, not independently reproduced. Training is resource-intensive and relies on a hand-crafted procedural task space. Sudden morphology changes can cause falls or failed adaptation; memory across attempts is not uninterrupted fault tolerance or deployment safety. Manipulation and general whole-body task capabilities are not established. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.
