Boston Dynamics picks Alexa's Prasad as CEO

Today: Boston Dynamics hands the chief executive job to an AI product veteran, TwelveLabs adds first-person video understanding, two papers put numbers on VLA routing and recovery, and open-source ROS 2 work lands on dialogue and dual-arm welding.

Boston Dynamics names Rohit Prasad as CEO

Boston Dynamics named Rohit Prasad as chief executive on 6 October, effective the next day. He replaces Amanda McMaster, who had been interim chief executive since Robert Playter stepped down in February, and is the third chief executive in the company’s history (The Robot Report).

Prasad spent 12 years at Amazon, helping build Alexa and later leading the Nova foundation model family, and almost 14 years before that at Raytheon BBN Technologies. Boston Dynamics frames the hire as accelerating its physical AI strategy, and Hyundai chair Jaehoon Chang points to robotics, AI product experience and Hyundai manufacturing.

The signal is where the company thinks the remaining work sits: a humanoid maker staffed with mechanical and controls talent hired someone who ships large learned systems to consumers, just as Atlas is integrated into Hyundai’s Metaplant America (opened in September) with a 13 degree-of-freedom hand built for simulation and reinforcement learning. Expect the differentiation to move from hardware to the learned layers developers build on.

TwelveLabs Pegasus 1.6 adds first-person video

TwelveLabs released Pegasus 1.6 on 6 October. The video understanding model now handles egocentric footage, video shot from the point of view of whoever is acting, which the company says it supported poorly before. It also improves recognition of people and objects in a scene, and reads still images natively through the same API and prompts (TwelveLabs).

The egocentric point is the robotics one. Teleoperation recordings and first-person action data are the raw material for imitation learning, and labelling that footage has mostly meant custom tooling. A model that turns unstructured first-person video into searchable text cheaply replaces a hand-built annotation pass, attacking the robot data bottleneck from the collection side. TwelveLabs published no benchmarks, so test it on your own teleoperation set first.

SWAP routes between VLA policies at every step

A paper on arXiv from 3 October proposes StepWise Action Policy Routing (SWAP), which composes several vision-language-action policies during execution instead of committing to one for a whole episode. Routing is cast as an offline reinforcement learning problem, and a learned critic selects the best policy at each decision step. It is under review for ICRA 2027, by Mousumi Das, Aditeya Prajapati, Abrar Anwar and Jesse Thomason (arXiv:2610.06926).

On real-world DROID tasks and LIBERO simulation, SWAP reports absolute improvements in real-world success up to 33 percentage points over fixed-policy execution and routing baselines, with successful trajectories 28.3 percent shorter in action steps. Step count matters as much as success rate in production, where every step is time and wear.

The premise is familiar: one checkpoint wins in some task states and loses in others, the problem behind the conditioning arguments in world-action models. SWAP keeps the policies frozen and learns only the switch.

Does a learned corrector beat a simple retreat?

A second paper from 3 October asks what most recovery work skips: is the improvement bigger than the base policy’s own run-to-run variation, and is the learned machinery better than something trivial? The authors evaluate a frozen pi-0.5 policy on four RoboTwin tasks over 3,888 paired episodes, with a same-seed re-run of the base policy as a placebo (arXiv:2610.06921).

The full pipeline raises success on one task by 13.5 percentage points (95 percent interval 9.4 to 17.7) and shows no detectable gain on the other three. Among failed base episodes on the responsive task, 43.2 percent succeed on a plain re-run, so many nominal rescues are just base policy variability. A fixed-time trigger plus a scripted return to an earlier joint configuration matched the learned pipeline’s net gain across two rounds, though not consistently by the authors’ own equivalence criterion. Anyone building recovery on a frozen model should measure against a placebo and a dumb baseline first.

DialoStack keeps the LLM out of the ROS 2 control flow

A developer released DialoStack under MIT, a set of ROS 2 packages for structured spoken dialogue, such as clinic intake, explaining a topic and checking comprehension, or a short quiz. An LLM wired straight into the speech loop drifts off task, re-asks things it already has and never decides the conversation is over, so the flow is plain Python state machine logic (states, turn limits, confirmation) and the model only interprets the utterance and phrases the reply. It began as a bachelor’s thesis, presented at WAF 2026 (ROS Discourse).

One action, /dialog/execute_task, drives everything and returns a confirmed slot frame as JSON. The stack ships a dialog manager with slot filling, explanation and quiz strategies, a fully local LLM path through Gemini or Ollama, speech recognition on faster-whisper and Silero VAD, Piper for streaming synthesis, and a BehaviorTree.CPP leaf node. It is a clean example of the action pattern keeping a non-deterministic model out of the layer that decides when a task ends.

hold_and_weld v0.3.0 for dual-arm welding

Another open-source release, hold_and_weld, targets dual-arm welding: one arm holds the workpiece while the second welds it, removing the need for dedicated fixtures. Version 0.3.0 targets ROS 2 Jazzy on Ubuntu 24.04 as a simulation proof of concept, with the grasping and welding halves usable independently (ROS Discourse).

Constraint-aware grasp sampling generates antipodal candidates on CAD surfaces with exclusion volumes that keep grasps away from weld seams and screw holes, filtered against gripper geometry and collisions, accepting STEP and URDF inputs. A weld finder extracts seams from CAD and mesh geometry as JSON weld paths. A configuration finder picks the approach joint configuration before OMPL plans to it with a custom Ceres inverse kinematics solver, rejecting candidates on IK failure, joint limit violations, low Yoshikawa manipulability or excessive joint steps. Removing fixtures is what makes cobot arms economically interesting; read it alongside how MoveIt 2 is being folded into vendor stacks.

FireDome points autonomous launchers at wildfire

FireDome, a California startup founded by Gadi Benjamini and Dr Adi Naor Pomerantz in 2024, is building a ground-based autonomous launcher that fires projectiles containing water or fire retardant up to 1,000 ft (about 305 m), covering roughly 70 acres (28.3 hectares) per unit (The Robot Report).

The constraint is honest. Covering wildland would need 1 to 1.5 mi (1.6 to 2.4 km) of range, so the company protects specific properties and infrastructure instead, outside code-mandated indoor suppression. The target is embers: Benjamini says 85 to 90 percent of structure losses in the wildland-urban interface in recent California fires came from wind-driven embers, not flame fronts. One demo system exists in Yuba County, California, with production units planned for spring 2027 and the launchers meant to appear inside incident command as a resource a fire boss controls. Handing authority to the on-scene commander is a useful precedent for supervision and intervention in safety-critical field robots.

Quick Bites

• Overview AI launched the OV Spark and OV Spark Pro inspection cameras with an onboard agent called Sparky. Each camera carries 10 inspection tools in one recipe, seven of them deep learning models for triggering, alignment, classification, segmentation, counting, OCR and anomaly detection, with recipes, inference and PLC communication running on the camera (The Robot Report).
• Tesla extended its Austin robotaxi hours to 11 pm from 10 pm, and Elon Musk said the main thing to solve is not running over pets that are hard to see at night, describing grey kittens on grey tarmac: a low-light, low-contrast detection problem (Electrek).
• Slides and talk videos from ROSCon JP 2026, held 4 and 5 August in Tsukuba, are now public (Open Robotics Discourse).
• Motion hardware: Nidec Drives extended its Commander S V2 AC drive range to 11 kW (15 hp), and Copley Controls released the R47 and R48 NanoPlus Micro Module servo drives, hardened for humidity, vibration and shock (Design World, Nidec, Design World, Copley).