New Technology / Robotics
Track robotics trends, industrial automation, machine intelligence and commercial deployment signals through curated technology summaries.
Figure 03 Humanoid Robot Learns 8 New Autonomous AI Skills (AI NEWS)
Topic
Humanoid Robot Autonomous Skills
Key insights
- Figure 03 humanoid robot showcases eight autonomous cleaning skills, including coordinated tool use and dynamic object handling, enhancing real-world cleaning efficiency
- Figured has demonstrated a humanoid robot with eight autonomous cleaning skills, showcasing advanced capabilities such as coordinated tool use and dynamic object handling. The robot's Helix zero two brain integrates three systems to enhance its operational efficiency and adaptability.
- Physical Intelligence deployed its Pi 0.5 model on an excavator, marking a first for heavy equipment outside lab settings, addressing real-world data challenges
- Data was collected using monocular cameras to map joystick movements to a 40 action space, enabling breakdown of tasks like trench digging into repeatable sub-tasks
- Joystick control works for VLA despite being nonlinear, aiming to create a generalized policy across multiple machine types and tasks
- Anthropics new code review tool uses multiple AI agents per pull request to identify bugs in parallel, increasing comment rates from 16% to 54%
Perspectives
Focus on humanoid robot capabilities and challenges in AGI development.
Humanoid Robot Capabilities
- Demonstrates eight new autonomous cleaning abilities
- Executes coordinated tool use with spray bottle and towel
- Implements dynamic object handling by repositioning tools
- Performs bimanual manipulation for efficient task execution
- Utilizes whole-body strategies for enhanced efficiency
- Throws objects dynamically without new algorithms
Challenges in Robot AGI Development
- Faces barriers in real-world machine data scarcity
- Struggles with interfacing hardware with industrial equipment
- Encounters difficulties adapting control from joint encoders to joystick inputs
- Relies on monocular cameras for data collection, limiting depth perception
- Assumes joystick control can generalize across various machine types
Neutral / Shared
- Integrates a three-part robot brain for operational efficiency
- Utilizes a vision-language-action model for task execution
- Implements a code review tool that improves bug detection efficiency
Metrics
payload capacity
20 kilograms kg
maximum weight the robot can carry
This capacity indicates the robot's potential for handling various cleaning tasks.
it has a payload capacity of 20 kilograms or about 33 pounds.
speed
1.2 meters per second m/s
movement speed of the robot
Speed is crucial for efficiency in cleaning operations.
with a speed of about 1.2 meters per second.
other
40 action space actions
the number of actions the joystick movements were mapped to
This indicates the complexity of tasks the model can handle.
mapped them to a 40 action space
other
84%
large pull requests flagged by the tool
This indicates the tool's effectiveness in identifying issues in substantial code changes.
large pull requests being flagged at 84% of the time
other
7.5 issues
average number of issues found per pull request
This metric highlights the tool's capability to detect multiple issues efficiently.
averaging 7.5 issues
Key entities
Timeline highlights
00:00–05:00
Figured has demonstrated a humanoid robot with eight autonomous cleaning skills, showcasing advanced capabilities such as coordinated tool use and dynamic object handling. The robot's Helix zero two brain integrates three systems to enhance its operational efficiency and adaptability.
- Figure 03 humanoid robot showcases eight autonomous cleaning skills, including coordinated tool use and dynamic object handling, enhancing real-world cleaning efficiency
05:00–10:00
Physical Intelligence has successfully deployed its Pi 0.5 model on an excavator, marking a significant advancement in heavy equipment automation. Anthropic's new code review tool has improved bug detection efficiency, significantly increasing comment rates on pull requests.
- Physical Intelligence deployed its Pi 0.5 model on an excavator, marking a first for heavy equipment outside lab settings, addressing real-world data challenges
- Data was collected using monocular cameras to map joystick movements to a 40 action space, enabling breakdown of tasks like trench digging into repeatable sub-tasks
- Joystick control works for VLA despite being nonlinear, aiming to create a generalized policy across multiple machine types and tasks
- Anthropics new code review tool uses multiple AI agents per pull request to identify bugs in parallel, increasing comment rates from 16% to 54%
- The code review process costs $15 to $25 per pull request and caught a critical change that could have broken production
- OpenAIs acquisition of OpenClaw in 2026 led to its loss of independent status, prompting Nvidia to prepare NemoClaw as an open-source alternative