Baidu Apollo
Baidu's Apache 2.0 autonomous driving platform (阿波罗). Version 11.0, January 2026, aims squarely at sweepers, delivery pods, patrol vehicles and campus shuttles rather than robotaxis.
Overview
Apollo has been open since 2017 and remains the most complete autonomous driving stack anyone gives away. Localisation, perception, prediction, planning, decision, control, mapping, plus CyberRT, the in-house middleware that replaced ROS for the real-time path. The engineering is C++ and the architecture assumes you are building a vehicle, not a research demo, which is both its strength and the reason a first build takes real effort.
The 11.0 release on 30 January 2026 is the interesting one. Rather than chasing full urban autonomy, it targets what Baidu calls functional vehicles: package delivery, street sweeping, security patrol and campus shuttles. The feature list reads accordingly, with hill start and stop, edge-hugging for kerbside sweeping, gate recognition for moving between public roads and enclosed campuses and automatic recovery when something goes wrong unattended. Localisation was rebuilt around fusing RTK, SLAM, vision and wheel odometry to hold centimetre accuracy where GPS does not reach, which is exactly the failure mode that kills logistics-park deployments.
Two caveats deserve emphasis. First, versioning: the GitHub tag list stops at apollo-v9.0.0 from 2023 while 10.0 and 11.0 are announced through the Apollo Open Platform and land on master, so treat the tags as historical and pin a commit. Second, the split between open code and cloud services keeps widening. Perception improvements in 10.0 and 11.0 lean on Baidu's platform for incremental training, and the repo carries an explicit research-purpose note on models and methods. For a team building a low-speed commercial vehicle in China, the stack is still the fastest route from nothing to a moving prototype, provided you plan around a Ubuntu-pinned toolchain and Chinese-language support channels.
Key Features
- ✓ Apache 2.0, 26,807 stars and 19,365 commits on master, last pushed 16 April 2026. One of the largest genuinely open autonomy codebases in existence.
- ✓ Apollo 11.0, launched 30 January 2026, targets functional vehicles: parcel delivery, street sweeping, security patrol and campus shuttle, with hill start and stop, edge-hugging for curb work and automatic recovery from abnormal states.
- ✓ CyberRT middleware with zero-copy messaging. Baidu claims a tenfold transport improvement and L4 capability on a single NVIDIA Orin from the 10.0 generation onward.
- ✓ BEV plus occupancy network perception with incremental training on your own data, and centimetre localisation fusing RTK, SLAM, vision and wheel odometry for GPS-denied yards and enclosed campuses.
- ✓ Full module set in C++: localisation, perception, prediction, planning, decision, control and mapping, with Dreamview Plus for visualisation and package-based module development since 9.0.
- ✓ Runs on Ubuntu 18.04, 20.04 and 22.04 with CUDA 11.8, NVIDIA Turing or newer including the 40-series, AMD GFX9/RDNA/CDNA under ROCm 5.1 and ARM builds for Orin.
Where it holds
- • A complete, industrially exercised autonomy stack under a permissive licence, which is not available from any Western equivalent at this scale.
- • Hardware breadth is unusual: NVIDIA, AMD under ROCm and ARM/Orin targets are all supported, so you are not locked to one vendor.
- • The 11.0 direction, low-speed functional vehicles, is where deployable autonomy money actually is right now, and the release reflects real fleet experience rather than a demo.
Where it breaks
- • Version numbering is confusing. The newest GitHub tag is still apollo-v9.0.0 while 10.0 and 11.0 arrive through the Apollo Open Platform and the master branch, so pin a commit rather than trusting the release list.
- • The build is heavy. Docker images run to tens of gigabytes and the Ubuntu and CUDA matrix is unforgiving, so budget a day before anything moves.
- • Newer perception models are increasingly tied to Baidu's cloud training platform rather than published as free weights, which thins out what you actually get offline.
- • English documentation trails the Chinese by a long way and the developer community is Chinese-first, so real troubleshooting happens on Chinese forums.
- • The repository states research purpose for the models and methods, so a commercial deployment needs its own safety case and probably a conversation with Baidu.
My Take
Apollo spent this cycle walking away from the robotaxi framing. Version 11.0, out on 30 January 2026, aims at sweepers, delivery pods, patrol vehicles and campus shuttles, which is honest about where deployable autonomy revenue sits in China today. The code stays Apache 2.0 with 26.8k stars and commits into April 2026, but the newest GitHub tag is still v9.0.0 because 10.0 and 11.0 ship through the Open Platform, so pin a commit and not a release. Clear a day for the first build: CUDA 11.8, a narrow Ubuntu matrix and Docker images measured in tens of gigabytes.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Free under Apache 2.0. The repository note states that models and methodologies are for research purposes and that commercial use should carry sufficient safety mechanisms. Production hardware kits, HD map data and the cloud simulation service are separate commercial arrangements with Baidu.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Baidu Apollo holds and where it breaks. The fastest check is your own workload.
Visit website ↗