Nemotron edge models and Digit 5 charging

Yesterday’s roundup covered a benchmark that frontier models fail. Today: open vision models sized for robot hardware, the battery supplier behind Digit 5, a 400 million dollar option on surgical robotics, three layers that carry the action in a VLA, and ROS 2 telemetry that can finally see Python nodes.

Nemotron 3.5 splits the vision model into server and robot tiers

NVIDIA released Nemotron 3.5 Super VL on 15 October, an open 120 billion parameter multimodal model, alongside two edge models in the same family (NVIDIA). Super VL and Nemotron 3.5 Bolt 74B-A9B, a 74 billion parameter mixture of experts with roughly 9 billion active, are available now, with the 4 gigabyte Bolt E4B following on 23 October.

The tiering is the part that matters for robotics. NVIDIA says Bolt E4B is built for physical and conversational agents on GeForce RTX, Jetson and RTX or DGX Spark, which puts the language and vision layer on the board next to the motors instead of round tripping to a server, the trade off we compared for Orin Nano and Orin NX.

The capability numbers are NVIDIA’s own: 5 times faster throughput than the next model in its class on the Artificial Analysis Intelligence Index leaderboard, and accuracy said to be comparable with NVIDIA’s largest reasoning model at roughly one fifth of the parameter count. The vision work targets OCR, tables, charts and structured documents, so think inspection reports rather than contact rich manipulation, and post training on your own data is allowed.

Agility names the battery supplier behind Digit 5’s charge time

Nyobolt is the battery technology provider for Digit 5, Agility Robotics said on 15 October (Agility). The pack figures were already public at the September launch: 90 minutes of runtime, a nine minute charge, a 10 to 1 run to charge ratio against Digit 4’s 2 to 1, and Agility’s claim of more than 20 hours of productive work in a 24 hour day. What is new is the supplier and the mechanism behind those numbers.

Nyobolt, founded in 2019 in Cambridge, UK, attributes its charge rate to proprietary anode materials, cell design and integrated power electronics rather than a larger pack, which is what counts when a humanoid has a fixed mass budget. The Robot Report reports it reached a 1 billion dollar valuation in May 2026.

A 10 to 1 ratio turns charging into fleet scheduling: one robot covers more than two shifts if the nine minute window is guaranteed, so the charge dock becomes part of the uptime calculation rather than a break in it. Agility says Digit has accumulated more than 65,000 hours of real world operation, and early access to Digit 5 with the Nyobolt system is expected in the first half of 2027. Our September roundup covers the rest of the Digit 5 hardware.

Stryker takes a 400 million dollar option on soft tissue robotics

Stryker will invest approximately 400 million dollars, 300 million pounds, in CMR Surgical as preferred equity, with an exclusive option to acquire the remaining equity and an exclusive United States sales agency agreement for the Versius portfolio (Stryker, CMR Surgical). Announced on 12 October and subject to regulatory approval, the deal leaves CMR leading business outside the United States.

Versius is a modular soft tissue system built to fit existing operating rooms rather than a dedicated suite. The Robot Report reports adoption in more than 30 countries and 45,000 cases, FDA 510(k) clearance for Versius Plus in cholecystectomy, and a pending 510(k) submission for benign gynecological indications. CMR has also been building an AI stack with NVIDIA, including IGX Thor, according to its own news page.

Surgical robotics rewards nearly the opposite of what general manipulation rewards: bounded workspaces, hard force limits, sterilization cycles, and verification paths a regulator accepts. Capital concentrating behind a few platforms sets the compliance bar every smaller manipulation company will be measured against.

Three of thirty six layers carry the action in MolmoBot

A mechanistic study of three frontier vision language action models finds that causally relevant information reaches the action expert through a narrow band of mid to late cross attention layers, which the authors call the action conditioning interface (arXiv:2610.20750).

Activation patching across pi 0.5, MolmoAct2 and MolmoBot shows that early to mid layer representations of the target noun token determine which object is selected, while mid to late layers carry the spatial information in text tokens that redirects the movement when patched. Where that spatial information lives depends on architecture: MolmoAct2 forms task conditioned representations at the text positions following the instruction, while pi 0.5 also stores them at image token positions.

The efficiency result is concrete: restricting MolmoBot’s action expert conditioning to 3 of 36 layers cut inference FLOPs by up to 33 percent, with task success at 46.7 percent against 46.9 percent for the default model across 3,000 evaluation episodes, and the authors report the same principle on a real robot deployment of pi 0.5. Read alongside the VLA primer and world action model conditioning, this is interpretability that pays for itself wherever inference latency and compute are budgeted.

ROS 2 telemetry learns to count Python nodes

Version 0.6.0 of ros2_pulse extends its LD_PRELOAD probe, which reads tracetools tracepoints rather than subscribing to topics, to cover the receive side (ROS Discourse). Python subscribers used to be invisible because rclpy is not instrumented; the probe now hooks the rmw_take tracepoint so rclpy and rclc receives are counted, while rclcpp still counts at callback_start so nothing is counted twice. Loaned messages get their own rate, and from Jazzy Fast DDS loans plain types by default, so a Float64 publisher may already be doing that without anyone asking. pulse-export serves the windows as Prometheus metrics on port 9464, with a Grafana and Docker Compose example in the repository, and packages cover Humble, Jazzy and Kilted.

Silent failures in a running stack are usually timing and QoS problems, which is why rosbag2 plus a live rate dashboard beats log spelunking after the fact.

In the same week, an RFC opened for gz_ros2_control to consolidate its integration test architecture: assert physical feedback such as joint position and effort responses instead of checking that processes and topics exist, share Gazebo sessions across assertions to cut CI runtime, and keep semantics aligned across Rolling down to Humble (ROS Discourse).

A teleoperated robot becomes a police case

Cologne police seized a 120 cm robot and two phones and opened an investigation into its 16 year old owner and a 19 year old companion, after a 16 year old girl reported that the robot grabbed her and made thrusting motions against her before she screamed and moved away (BBC). Police say the pair may have streamed the incident live and are still establishing whether the movements were deliberately controlled.

Nothing here is proven beyond the report, and the case is an ordinary criminal investigation rather than an autonomy failure. The narrow technical reading: this was remote operation by people, so liability sits with the operator, and machines cheap enough to walk up behind a stranger in a street have arrived before any shared convention about operating one responsibly.

Quick Bites

• Block sparse featurizers applied to VLA models recover multi dimensional features for perception, language, proprioceptive state and episode phase, and feature based data attribution isolates fewer than 1.4 percent of training tokens whose masking removes a targeted behaviour such as failure recovery (arXiv:2610.20767).
• Boston Dynamics’ Marco da Silva says customers increasingly start with programs rather than single robots, describing early majority behaviour, with Spot running thermal, visual, air leak and vibration inspections, and deployment work covering power, data paths and workforce preparation (ASSEMBLY).
• The ROS Cloud Robotics working group meets on 19 October at 16:00 UTC to finish the first draft of its logging and observability guide (ROS Discourse).