NavDP (Navigation Diffusion Policy)
Robotics modelEnd-to-end mapless RGB-D navigation diffusion policy with a privileged-information-guided trajectory critic, trained purely in simulation for zero-shot sim-to-real transfer across wheeled, quadruped, and humanoid robots. [1] [3] [4]
Setup & usage →CheckpointsModel hub / homepageGetting started
- For
- Embodied navigation researchers evaluating sim-to-real diffusion policies in Isaac Sim / Isaac Lab or deploying on wheeled/legged robots under non-commercial terms [1]
- Parameters
- Not stated as a scalar in README or arXiv:2505.08712 abstract (navdp_pretrain.ckpt is 543 MB and… [1] [4] [5]
- Embodiments
- Wheeled mobile robots (Isaac Sim wheeled benchmark and LeKiwi hardware deployment guide via LoGoPlanner),… [1] [4] [5]
- Source rights
- Some licenses are unresolved. Includes noncommercial terms. Review licenses
Source researched · Not tested by us
Sources checked · Research details
Licenses & related hardware
- Ecosystem
- Isaac Sim
- License
- CC-BY-NC-SA-4.0 (declared in README; no root LICENSE file) (noncommercial), Conflicting (HF metadata: CC-BY-SA-4.0; gating access agreement: CC-BY-NC-SA-4.0) (noncommercial), CC-BY-NC-SA-4.0 (declared in README) (noncommercial) · license not fully assessed for model weights · Details by part
Overview
Navigation Diffusion Policy (InternRobotics/NavDP, ICRA 2026, arXiv:2505.08712) is an end-to-end mapless RGB-D navigation policy developed by Shanghai AI Laboratory, Tsinghua University, Zhejiang University, and The University of Hong Kong. Trained exclusively in simulation on over one million meters of navigation trajectories across 3,000 scenes, NavDP targets zero-shot sim-to-real transfer across point-goal, image-goal, and no-goal exploration tasks. At pinned commit bff5eb8857f50e082243fb455d312409daabc3aa, the repository ships the decoupled navdp_server.py policy server, an Isaac Sim 4.2.0 / Isaac Lab 1.2.0 evaluation benchmark, and related baselines including CoRL 2026 X-NavDP and LoGoPlanner. [1] [2] [4]
Model card summary
- Architecture
Unified Transformer diffusion policy and trajectory critic (NavDP_Policy in baselines/navdp/policy_network.py): encodes 224x224 RGB-D history (8-frame memory via NavDP_RGBD_Backbone using Depth Anything V2 representations) plus optional point, pixel, and image goal encoders into 384-dim tokens, fused by an 8-layer, 8-head causal Transformer decoder with a 10-step DDPM noise scheduler (action_head outputting 24-step 3D waypoint trajectories) and a privileged-information-guided critic_head.
- Base model
Depth Anything V2 (RGB foundation representation in NavDP_RGBD_Backbone)
README acknowledgements note Depth Anything V2 provides the foundation representation for RGB image observations.
- Parameter count
Not stated as a scalar in README or arXiv:2505.08712 abstract (navdp_pretrain.ckpt is 543 MB and x-navdp_posttrain.ckpt is 873 MB on Hugging Face)
Exact parameter count is not stated in README or abstract; inferred scale from verified Hugging Face checkpoint sizes.
- Input modalities
- rgbdrgbdepth
Local 224x224 RGB-D observation history (8 frames) conditioned on optional 3D point goal, 2D pixel goal, or RGB image goal.
- Output modalities
- waypoints
Predicts 24-step 3D local navigation waypoint trajectories (action_head dim 3) ranked by a scalar trajectory critic (critic_head) and tracked by an asynchronous MPC controller.
- Action space
24-step local 3D waypoint trajectory sequences generated via 10-step DDPM denoising and ranked by a learned trajectory critic, executed asynchronously by an external MPC or locomotion controller.
- Trained / supported embodiments
- Wheeled mobile robots (Isaac Sim wheeled benchmark and LeKiwi hardware deployment guide via LoGoPlanner)
- Cross-embodiment quadruped (Unitree Go2) and humanoid (Unitree G1) policies included in the X-NavDP Hugging Face release
- Training datasets
- Simulated navigation dataset spanning >1 million meters of navigation experience across 3,000 scenes (InternRobotics/InternData-N1 on Hugging Face, >1T scale, CC-BY-SA-4.0, auto-gated)
- Evaluation scenes from InternRobotics/Scene-N1 (Cluttered-Easy, Cluttered-Hard, InternScenes-Home, InternScenes-Commercial; CC-BY-SA-4.0, auto-gated)
- Interfaces
- Decoupled HTTP inference server (baselines/navdp/navdp_server.py)
- Isaac Sim / Isaac Lab evaluation and keyboard teleoperation CLIs (eval_*_wheeled.py, teleop_*_wheeled.py)
Checkpoints
| Checkpoint | Format | Size | Access | Pinned revision |
|---|---|---|---|---|
| NavDP InternVLA-N1 System-1 latest checkpoint (request-only via Google Form) | pytorch | Not stated | request-only | master / bff5eb8857f5 |
| InternRobotics/X-NavDP (navdp_pretrain.ckpt + x-navdp_posttrain.ckpt + Go2/G1 policies) | pytorch | 1.32 GB | open | 7cee38a8d830 |
Primary NavDP System-1 checkpoint link is a Google Form (HTTP 200 verified); the related X-NavDP repo linked in the README openly hosts navdp_pretrain.ckpt (543 MB), x-navdp_posttrain.ckpt (873 MB), and Go2/G1 policy.pt files (total 1.32 GiB). LoGoPlanner (4.339 GiB, auto-gated) is also linked as a separate baseline.
Is this for you?
- Use cases
- Mapless point-goal, image-goal, pixel-goal, and no-goal exploration navigation via decoupled HTTP server (navdp_server.py)
- Asynchronous closed-loop System-1 navigation benchmarking in Isaac Sim 4.2.0 / Isaac Lab 1.2.0 with MPC trajectory tracking across cluttered, home, and commercial scenes
- Cross-embodiment RL post-training via X-NavDP for wheeled, quadruped (Unitree Go2), and humanoid (Unitree G1) platforms
- Known limitations
- No standalone root LICENSE file; README declares CC-BY-NC-SA-4.0 for open-sourced code (non-commercial, share-alike), while InternRobotics/X-NavDP on Hugging Face has no declared license metadata.
- The latest InternVLA-N1 System-1 NavDP checkpoint is request-only via Google Form, whereas the X-NavDP Hugging Face repo openly hosts navdp_pretrain.ckpt and x-navdp_posttrain.ckpt.
- Full closed-loop simulation evaluation requires Isaac Sim 4.2.0, Isaac Lab 1.2.0, and auto-gated Hugging Face scene assets (InternRobotics/Scene-N1).
- Outputs local waypoint trajectories that require an external MPC or low-level locomotion controller for physical execution.
- Safety and limitations
- CC-BY-NC-SA-4.0 notice in README prohibits commercial use and requires share-alike redistribution; the X-NavDP Hugging Face repo lacks explicit license metadata.
- NavDP predicts 24-step local waypoint trajectories and critic scores, relying on an asynchronous MPC or low-level robot controller for collision-safe physical actuation.
- Trained purely in simulation; zero-shot sim-to-real transfer performance depends on RGB-D camera calibration, depth noise, and robot dynamics.
- Open Source Robotics has not submitted the checkpoint form, downloaded Hugging Face weights, installed Isaac Sim 4.2, run inference, or tested NavDP on physical robots.
Compute & fine-tuning
- Inference compute
- Documented hardware: NVIDIA CUDA GPU (NavDP_Policy defaults to cuda:0; Isaac Sim 4.2.0 / Isaac Lab 1.2.0 required for simulation benchmark; exact VRAM minimum not stated)
- Precision:
fp32
cuda:0 default and float32 image preprocessing verified in baselines/navdp/policy_network.py and policy_agent.py; scalar VRAM minimum and latency are omitted.
- Fine-tuning
- Supported: Yes
- Methods: RL post-training across heterogeneous embodiments via X-NavDP (baselines/x-navdp, CoRL 2026)
- Recipe: Upstream fine-tuning documentation
Main NavDP README focuses on server inference and Isaac Sim evaluation; RL post-training is provided via the X-NavDP baseline.
- Host computer
Linux host with NVIDIA CUDA GPU for the NavDP inference server and Isaac Sim 4.2.0 / Isaac Lab 1.2.0 simulation benchmark; exact VRAM minimums are not stated.
Inferred from cuda:0 policy default and Isaac Sim 4.2.0 requirements.
- Operating system
- Linux
Isaac Lab ./isaaclab.sh installation and bash/conda workflows target Linux.
[1] · Editorial assessment
- Dependencies
- Python 3.10, PyTorch, diffusers (DDPMScheduler), OpenCV, Depth Anything V2 components
- NVIDIA Isaac Sim 4.2.0.2 and Isaac Lab 1.2.0 (for closed-loop evaluation benchmark)
Setup & usage
Open the upstream getting-started guide [1] · Model card / documentation [1]
- Create a Python 3.10 conda environment (navdp) and install baselines/navdp/requirements.txt.
- Obtain a checkpoint via the Google Form (or download navdp_pretrain.ckpt / x-navdp_posttrain.ckpt from InternRobotics/X-NavDP) and launch python navdp_server.py --port 8888 --checkpoint <path>.
- For closed-loop simulation evaluation, create a separate Python 3.10 conda environment with Isaac Sim 4.2.0.2 and Isaac Lab 1.2.0 and install root requirements.txt.
- Download InternRobotics/Scene-N1 evaluation scenes from Hugging Face and run eval_pointgoal_wheeled.py, eval_imagegoal_wheeled.py, or eval_nogoal_wheeled.py.
Upstream-reported benchmarks
Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.
| Benchmark | Upstream-reported score | Scope & caveats |
|---|---|---|
| Simulated (InternVLA-N1 System-1 / Isaac Sim) and Real-World Zero-Shot Navigation (ICRA 2026, arXiv:2505.08712) | Trained purely in simulation (>1M meters across 3,000 scenes), NavDP is reported to significantly outperform prior state-of-the-art baselines in both simulated and real-world navigation with zero-shot sim-to-real transfer across diverse environments and embodiments. | Upstream-reported qualitative/comparative summary in arXiv:2505.08712 abstract (numeric score tables are not reproduced in the repository README); OSR has not run Isaac Sim or hardware evaluations. |
Project notes
Architecture, checkpoints, and Isaac Sim benchmark
In baselines/navdp/policy_network.py, NavDP_Policy encodes 8-frame histories of 224×224 RGB-D observations (leveraging Depth Anything V2 foundation representations) alongside optional point-goal, pixel-goal, and image-goal encoders into 384-dimensional tokens. An 8-layer, 8-head causal Transformer decoder with a 10-step DDPM noise scheduler denoises 24-step 3D local waypoint trajectories (action_head), while a paired critic_head trained with privileged simulation supervision scores candidate trajectories to avoid collisions. During closed-loop evaluation, navdp_server.py serves waypoint predictions over HTTP to an asynchronous MPC trajectory tracker in Isaac Sim 4.2.0 / Isaac Lab 1.2.0. [1] [3]
Checkpoint access is split across two paths. The README links a Google Form (verified HTTP 200) to request the latest InternVLA-N1 System-1 NavDP checkpoint. Separately, the README's baseline table links InternRobotics/X-NavDP on Hugging Face (ungated, revision 7cee38a8d8308874d2b8488783c612f42060ac41, 1.32 GiB total), which openly hosts navdp_pretrain.ckpt (543 MB), x-navdp_posttrain.ckpt (873 MB), and Unitree Go2 / G1 low-level policies. Exact parameter counts, minimum GPU VRAM, and server decision frequency (Hz) are not stated in the README. [1] [5] [6]
Upstream-reported navigation performance
In arXiv:2505.08712, the authors report that simulation-only training with privileged-information critic guidance enables NavDP to significantly outperform prior state-of-the-art navigation baselines in both simulated and real-world zero-shot evaluations across diverse environments and embodiments. Numeric benchmark tables are not included in the repository README or paper abstract; all performance claims are upstream-reported, and Open Source Robotics has not submitted the checkpoint form, downloaded Hugging Face weights, installed Isaac Sim 4.2, run training, fine-tuning, or inference, or tested NavDP on physical robots. [1] [4]
What we checked
Source researched only; no installation, training, fine-tuning, inference, benchmark, or physical robot test performed. No affiliate links or reused media.
Not yet established
- Documented closed-loop control frequency (Hz)
Updates
- Upstream last updated
- Sep 7, 2026 [2]
- Latest release
- Not established
- Profile updated
- Nov 21, 2026
- Sources last checked
- Nov 21, 2026
- Next review
- Feb 19, 2027
- Reviewed revision
- commit snapshot; no GitHub release / bff5eb8857f50e082243fb455d312409daabc3aa · software repository
Separate license matrix
Include NavDP as an active sim-to-real navigation diffusion policy with inspectable PyTorch code, CC-BY-SA-4.0 Hugging Face datasets, and open X-NavDP pretrained/posttrained weights, while noting CC-BY-NC-SA-4.0 code terms and the request-only form for the latest System-1 checkpoint.
| Layer | Availability / license |
|---|---|
| Code | available · CC-BY-NC-SA-4.0 (declared in README; no root LICENSE file) README states open-sourced code is under CC-BY-NC-SA-4.0 (non-commercial, share-alike); no standalone root LICENSE file is present. [1] [2] |
| Model weights | available · License not fully assessed Latest InternVLA-N1 System-1 NavDP checkpoint is request-only via Google Form (HTTP 200 verified); InternRobotics/X-NavDP openly hosts navdp_pretrain.ckpt and x-navdp_posttrain.ckpt (1.32 GiB total, ungated) without declared HF license metadata. [1] [5] [6] |
| Training datasets | available · Conflicting (HF metadata: CC-BY-SA-4.0; gating access agreement: CC-BY-NC-SA-4.0) InternRobotics/InternData-N1 (>1T scale) and InternRobotics/Scene-N1 are auto-gated on Hugging Face with conflicting license declarations: HF metadata tags list cc-by-sa-4.0 while their access agreements state CC BY-NC-SA 4.0; maintainer clarification is required before relying on commercial reuse. [1] [6] |
| Documentation | available · CC-BY-NC-SA-4.0 (declared in README) Comprehensive repository README covering NavDP server, Isaac Sim benchmark, baselines, and LeKiwi deployment. [1] |
License split across code, weights, and datasets
The repository contains no standalone root LICENSE file; the README declares that open-sourced code is licensed under CC-BY-NC-SA-4.0 (non-commercial, share-alike). On Hugging Face, InternRobotics/X-NavDP carries no declared license metadata in its card. For the InternRobotics/InternData-N1 and InternRobotics/Scene-N1 datasets, both repositories are auto-gated and contain conflicting license declarations: their Hugging Face metadata tags list CC-BY-SA-4.0, whereas their gating access agreements state CC-BY-NC-SA-4.0. Because this conflict is unresolved in the upstream materials, adopters should obtain clarification directly from the maintainers before relying on commercial reuse permissions for either dataset. [1] [2] [5] [6]
References
Each source links to the exact revision we read. “Used for” shows where this page relies on it.
- S001 NavDP repository README
NavDP HTTP server setup, Google Form link for InternVLA-N1 System-1 checkpoint, Isaac Sim 4.2 / Isaac Lab 1.2 benchmark, X-NavDP / LoGoPlanner baseline links, and CC-BY-NC-SA-4.0 code license.Revision, date & where cited
Shanghai AI Laboratory (InternRobotics) · github.com · revision
bff5eb8· Accessed Nov 21, 2026
Used for: Overview, Model card summary, Checkpoints, Is this for you? and 6 more - S002 GitHub repository metadata and pinned commit
Last commit 2026-09-07; no root LICENSE file; not archived; no GitHub releases; ICRA 2026 description.Revision, date & where cited
InternRobotics/NavDP / GitHub · github.com · revision
bff5eb8· master / · Accessed Nov 21, 2026
Used for: Overview, Is this for you?, What we checked, Separate license matrix - S003 NavDP model implementation (baselines/navdp/policy_network.py)
NavDP_Policy: 224x224 RGB-D 8-frame memory, point/pixel/image goal encoders, 8-layer 8-head 384-dim Transformer decoder, 10-step DDPMScheduler, 24-step 3D action_head, and critic_head.Revision, date & where cited
InternRobotics/NavDP contributors · github.com · revision
bff5eb8· Accessed Nov 21, 2026
Used for: Model card summary, Is this for you?, Compute & fine-tuning, Architecture, checkpoints, and Isaac Sim benchmark - S004 NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance (arXiv:2505.08712)
End-to-end simulation-trained Transformer diffusion policy and privileged-information-guided critic over >1M meters and 3,000 scenes; reports zero-shot sim-to-real transfer outperforming prior baselines.Revision, date & where cited
Shanghai AI Lab, Tsinghua, ZJU, HKU (ICRA 2026 / arXiv) · arxiv.org · arXiv:2505.08712v3 abstract · Accessed Nov 21, 2026
Used for: Overview, Model card summary, Is this for you?, Upstream-reported benchmarks and 1 more - S005 Hugging Face API metadata for InternRobotics/X-NavDP
Verified ungated repo (SHA 7cee38a8d8308874d2b8488783c612f42060ac41, no declared HF license) containing navdp_pretrain.ckpt (543 MB), x-navdp_posttrain.ckpt (873 MB), and Unitree G1 / Go2 policy.pt files (1.32 GiB total).Revision, date & where cited
InternRobotics / Hugging Face · huggingface.co · HF API blobs=true 2026-11-21 · Accessed Nov 21, 2026
Used for: Model card summary, Checkpoints, Is this for you?, Setup & usage and 2 more - S006 Hugging Face datasets (InternData-N1, Scene-N1) and NavDP checkpoint form reachability
Verified InternData-N1 and Scene-N1 (auto-gated, cc-by-sa-4.0) and HTTP 200 reachability of the NavDP System-1 Google Form.Revision, date & where cited
InternRobotics / Hugging Face · huggingface.co · HF API & HTTP GET 2026-11-21 · Accessed Nov 21, 2026
Used for: Model card summary, Checkpoints, Is this for you?, Setup & usage and 2 more
Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1
Change history
2026-11-21 — Initial 0.8.0-draft.1 model profile for NavDP (Transformer diffusion policy + critic, form-gated System-1 checkpoint + open X-NavDP weights, CC-BY-NC-SA-4.0 code).