Rohit Prasad Becomes Boston Dynamics CEO as Atlas Enters Factory Training

The appointment and its scope

Boston Dynamics announced Rohit Prasad as its chief executive on 6 October 2026, with the appointment taking effect on 7 October. A proposed seat on its board remains subject to approval. The company says his remit includes bringing AI and robotics into commercial use across industrial and enterprise settings. [1]

Prasad previously held senior vice president and head scientist roles for Alexa and Artificial General Intelligence at Amazon. Before Amazon, he spent nearly 14 years at Raytheon BBN Technologies working on machine learning. Boston Dynamics presents that product-development experience as a reason for the appointment. It announces no new Atlas controller, task score or robot delivery in this release. [1]

What his Amazon work involved

Amazon’s account of his career says Prasad joined in 2013 during Alexa’s early development and later led the creation of Nova and the AGI organization. Those are team-leadership roles; the models were developed by researchers and engineers across those organizations. [2]

A June 2022 Amazon Science interview identifies him as Alexa AI senior vice president and head scientist. It discusses connecting dialogue history, information retrieval and response generation. These systems had to interpret successive requests and choose a useful response from incomplete context. [3]

Amazon introduced Nova on 3 December 2024. The announced family included a text model, models accepting text, images and video, and separate image and video generators. This gives concrete context to his foundation-model experience. None of those input and output descriptions establishes an Atlas motion-control system. [4]

For a robot carrying a part, understanding the instruction leaves several physical decisions open. A controller must choose a grasp, check reach, move without losing support, and detect an unsuccessful placement. Prasad’s appointment provides no evidence that these decisions became more reliable on 7 October. That requires comparisons on the same tasks and hardware.

The Atlas record before he arrived

[5] [6] [1]

The September report describes placing automotive parts in assembly order. It also reports discussions with existing Spot and Stretch customers about further Atlas training in 2027. Those discussions do not confirm completed deployments at those customers. [6]

Factory learning had begun before the CEO appointment. Future progress can be judged against that starting point. A trained sequence becoming routine production work needs a record of acceptable parts delivered under the customer’s actual timing and supervision requirements.

For hardware generations and documented handling experiments, read the Atlas technical guide.

Count the work between human interventions

Start with a defined batch of parts, source containers and destination positions. Record every requested transfer from the work order through destination acceptance. A part placed in the wrong slot is a failed transfer even when the grasp succeeds. Keep first-attempt completion separate from completion after an automatic retry or a person’s correction.

Preserve the time spent waiting for a supervisor, resetting the robot or clearing a failed placement. A remote operator moving an arm, a worker repositioning a part and an engineer restarting software are different interventions. Count each category and its duration. A high completion rate can coexist with a large amount of assistance.

Repeat the same record across shifts and part changes. Otherwise, an easy batch or a technician stationed beside the robot can make two installations look comparable when their operating conditions differ.

Software connections and safe recovery need their own checks

Boston Dynamics says Orbit supports remote monitoring, task assignment and connections to manufacturing execution and warehouse management systems. Atlas can also use VR or tablet control. The company describes person detection that pauses the robot when someone enters a surrounding area, alongside padding and reduced pinch points. [7]

An integration trial should check what happens after a rejected placement, a lost work-order connection or a stopped robot. The factory system must retain a consistent record of which part moved and which job remains unfinished. Repeating an already completed command can create a second handling error.

A fleet update also needs a version record and repeat tests on existing jobs before broader rollout. Record changes in recovery behavior and operator workload. For shared work areas, assess stopping behavior with the intended load, approach direction and people nearby. The reviewed material does not give a numeric stopping-distance envelope for every workstation. [7]

Maintenance belongs in the cost calculation

The manufacturer says Atlas limbs can be replaced in the field in under five minutes. That is a component-replacement claim. The same account does not publish average fault frequency, diagnosis time or a complete cost per shift. [7]

A customer would need to add purchase or lease charges, spare parts, integration, operator training, supervision and lost production during repairs. Divide the cost over a stated period by accepted transfers in that period. This makes frequent retries, long stoppages and damaged parts visible in the result.

For the next Atlas deployment report, a named site and task would establish where the robot worked. Shift duration, accepted output and intervention logs would then show how much of that work the robot completed on its own. Those records would support a technical assessment of the company under Prasad’s leadership.