You can now buy a Cybercab and hail a driverless Toyota, and the simulation stack is keeping pace
Pony.ai and Toyota opened fully driverless service in Guangzhou, Tesla opened Cybercab orders, and training robots in simulation hit 60% real-world success.
- Pony.ai and Toyota launched fully driverless commercial services in Guangzhou with the Gen-7 bZ4X Robotaxi, aiming for around 1,000 vehicles across major Chinese cities, hail-able 24 hours a day. Self-reported m:…4533
- Tesla opened limited orders for personally-owned Cybercabs at a $30,000 one-time fee plus $300 a month, targeting a future monthly price of $115, according to the Trucks VC newsletter. m:…1871
- Scale AI acquired radar company Echodyne for $5.3 billion, and Zeromatter raised $160 million in a round led by Spark Capital, according to the Trucks VC newsletter. m:…1871
- Duality AI's Falcon 6.4 simulates 30 to 40 times faster than real time under specific configurations, and its Falcon Agent automates CAD ingestion and sensor tuning. Self-reported m:…8098
- In a preprint, PEAR trained open-weight 27-billion-parameter reasoning models for real robot control using only simulation, reaching 60% success on real dual-arm robots against 0% for the base model. Self-reported m:…0621
On the road
Pony.ai and Toyota have launched fully driverless commercial services in Guangzhou with the Gen-7 bZ4X Robotaxi, built from the ground up for Level 4 autonomous driving. Passengers in Guangzhou can hail the vehicle 24 hours a day, and the companies aim to deploy around 1,000 bZ4X robotaxis across major Chinese cities. The two companies have collaborated on multiple generations of robotaxis since 2019, and in 2023 Pony.ai, Toyota Motor China and GAC Toyota established a joint venture with an investment of more than RMB 1 billion. The first mass-produced bZ4X Robotaxi rolled off the line in February 2026, and the companies say they are developing the infrastructure to support fully driverless services at scale, including vehicle maintenance, charging and routine fleet servicing. m:…4533
On the other side of the Pacific, the model is ownership rather than hailing. Tesla opened limited orders for personally-owned Cybercabs with a $30,000 one-time fee and a $300-a-month subscription, and targets a future monthly price of $115 a month, according to the Trucks VC newsletter. May Mobility targets the deployment of over 10,000 autonomous ID. Buzz vehicles by 2031. The robotaxi is no longer a pilot project with a safety driver. It is a product with a price list, in two different business models at once. m:…1871
The money around it
The deals around the vehicles are moving just as fast. Scale AI acquired radar technology company Echodyne for $5.3 billion, Zeromatter raised $160 million in a funding round led by Spark Capital, and Path Robotics raised $300 million in a round led by Georgian, all according to the Trucks VC newsletter. Researchers estimate that autonomous trucking costs will match human-piloted per-mile costs in the US around 2028. The 2028 estimate is the one to watch: if trucking hits cost parity first, the fleet economics of autonomy get proven on freight before they get proven on passengers. m:…1871
The simulation layer
None of this ships without simulation, and the simulation tooling had its own release this week. Duality AI released Falcon 6.4, which achieves 30 to 40 times faster-than-real-time simulation under specific configurations, migrates to Unreal Engine 5.8, and makes raytracing the default mode. Its Falcon Agent automates CAD ingestion and sensor tuning, along with behavior scripting, and the release includes Cadaclysm for CAD geometry processing. The base system requirement is an NVIDIA RTX 4070 Ti or higher, which puts this class of simulation on a single workstation-class GPU rather than a cluster. m:…8098
The more important change is architectural. Rook replaces the deprecated Vibe Sim as a containerized world agent harness, enabling the deployment of predictable, neuro-symbolic agents, and DuCoSim introduces co-simulation with external physics engines including MATLAB, Ansys and Newton. Duality says Falcon 6.4 shifts the focus of digital twins to grounding AI agents for physical world reasoning. The twin is no longer a picture of the factory. It is the place where the agent learns how the physical world behaves. m:…8098
A preprint published this week shows how far simulation-only training has come. PEAR, a post-training engine for agentic robotics, trains open-weight 27-billion-parameter reasoning models for real robot control using only simulation-based reinforcement learning, with no human demonstrations. The trained model achieves a 60% success rate on real dual-arm robots, whereas the base model achieves 0%. On LIBERO-Pro it reached 57.4% success against 0.75% for pi_0.5, a demonstration-tuned model, and the authors report superior robustness to task changes compared with demonstration-tuned models. Sixty percent is not a finished product. It is a training method that works without the most expensive input, human demonstrations, and that changes who can afford to build a robot brain. m:…0621
What it means
Deployment and training are converging. Robotaxis are for sale and for hire, the simulation that trains physical agents runs 30 to 40 times faster than real time, and open-weight models can now learn robot control in simulation alone. The bottleneck shifts to the unglamorous parts Pony.ai and Toyota name explicitly: maintenance, charging and routine fleet servicing, and the cost per mile that decides whether the 2028 trucking parity estimate holds. The robots are ready to work. The question now is who can afford to keep them working. m:…4533 m:…0621 m:…1871 m:…8098
What is still unresolved
- How fast the Guangzhou robotaxi service scales beyond launch, and to which cities the 1,000-vehicle target extends first.
- Whether Tesla's $115-a-month Cybercab target is reachable, and what the $300-a-month subscription actually covers.
- Whether PEAR's simulation-only training transfers to robot types and tasks beyond the dual-arm setups tested.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.