← Read the Austin Robotics articleAustin Robotics · October 28, 2026tatbot.ai
Agentic Robotics
Hugo Ponte · Austin Robotics · October 28, 2026

Index

03Hugo07Live Demo09What Is an Agent?11Agents for Software14Agents for 3D18Agents for Robotic Control24System 1 + System 231Agents for Autonomous Research
My Name Is Hugo
Me Llamo Hugo
Je M’appelle Hugo

YouTube Streamer

youtube.com/@hu-po
Zero Data Self Play PretrainingWorld HarnessesRecursive Self-ImprovementWorld Models with AstraKernel RLZero-Shot LearnersTattoo RobotLanguage Guided RewardsAutoscaling InferenceScaling Synthetic DataAction ReasoningGPT5 Code
United States map with affiliation locations
hu-po

Robotics Researcher

Founder

@tatbot / YouTube
Serpent, 3RL on silicone skin▶ Watch original
silicone skin
Self portrait, forearm▶ Watch original
forearm tattoo
Evening calibration with new touch probe palette▶ Watch original
palette calibration
Recorded research footage · early human self-test does not establish readiness for use on other people.
Live Demo

Tatbot Hardware

Tatbot v6 Gaussian-splat capture from tatbot.ai
NVIDIA Jetson AGX Thor developer kit product imageNVIDIA DGX Spark product image
2xD405
Depth: 30fps | 1280x720
RGB: 30fps | 1280x720
Tattoo pen
Trossen WXAI arm
6 DoF | 1.5kg payload
Acc./Rep.: 1mm
Palette
PETG-CF | 6 inkcaps
D555
Depth: 30fps | 640x360
RGB: 30fps | 1280x720
Touch display
RPi camera
60fps | 1920x1080
Raspberry Pi
8GB | 4-core A76
Touch probe
Thor
128GB unified
14-core Arm CPU
Blackwell GPU
2070 TFLOPS (FP4)
Spark
128GB unified
20-core Arm CPU
Blackwell GPU
1000 TFLOPS (FP4)
What Is an Agent?

LLM + Harness

LLM
↔
Harness
Tools

Read, write and act

Context

Task, memory and evidence

Loop

Observe, decide and check

An action returns a result. The next decision uses it.
Agents for Software

Claude Code

Claude Code desktop — parallel agents product demo▶ Watch original
Anthropic · Desktop Demo (2026) · Claude Code

Replit Agent

Replit Agent — Miro brief to working app▶ Watch original
Replit · Miro to App Demo (2025)
Agents for 3D

MeshFormer

Direct 3D asset generation
Text or image
RGB + normals
Textured mesh
Sourdough bread
Sourdough bread
Golden goblet
Golden goblet
Chimpanzee with headphones
Chimpanzee with
headphones
Liu et al. (NeurIPS 2024) · Project + examples

Claude + Blender MCP

Claude + Blender MCP▶ Watch original
Siddharth Ahuja (2025) · Original Demo · Blender MCP

Tatbot Palette 🎨

Original five-cap palette
v1
Centered cap grid and three probe tips
v2
Palette v3 concept
v3
Palette v4 probe and camera concept
v4
Palette v5 elongated concept
v5
Palette v6 inline concept
v6
Palette v7 probe arc concept
v7
Palette v8 beam saddle concept
v8
Palette v9 fit review
v9
Palette v10 feedback revision
v10
Palette v11 redesign
v11
Palette v11 compact redesign
v12
Agents for Robotic Control
Research example · see original source for evaluation conditions.

Deep Visuomotor Policies

Images + Robot Configuration In, Motor Torques Out
Original Figure 2 from End-to-End Training of Deep Visuomotor Policies: convolutional visual features, spatial softmax and feature points combined with robot configuration to produce motor torques
Levine, Finn, Darrell & Abbeel · 2015 / JMLR 2016 · Figure 2
Research example · see original source for evaluation conditions.

Agents Controlling Robots

sensho · conveyor sorting simulation▶ Watch original
Conveyor sim · @sensho
Interlocked object manipulation example shared by Qineng Wang▶ Watch original
Interlocked parts · Qineng Wang
Robot painting example shared by thijs▶ Watch original
Painting · @cdngdev
Alex Nichol · Astra arranges a circle▶ Watch original
Circle of blocks · @unixpickle
Harrison Kinsley (@Sentdex) robot-control post (excerpt)▶ Watch original
Mobile robot · @Sentdex
Jay Chooi (@chooi_jeq) robot-control post (excerpt)▶ Watch original
Robot benchmark · Jay Chooi
Research example · see original source for evaluation conditions.

The Bitter Lesson Comes to Robotics

Robocurve original X video: score-doubling chart and Gemini 4 Argon versus GPT-5.6 Sol robot footage▶ Watch original
Robocurve original video (CC BY 4.0) · Paper (v2)
Research example · see original source for evaluation conditions.

Computer Use

OSWorld 2.0 offline score versus simulated time
Meta Muse
Official Meta Muse Mac screenshot organizing Downloads
Grok Bot
Official Grok Bot demo with CRM and outreach windows
OSWorld: partial task score, simulated time
Research example · see original source for evaluation conditions.

Pretraining and Post-Training Compute

01

Pretraining

Learn a starting model

02

Post-training

Adapt its behavior

03

Evaluation

Test the resulting behavior

Evaluation tells us whether a change helped.
Research example · see original source for evaluation conditions.
System 1 + System 2

Thinking, Fast and Slow

Original 2011 US hardcover cover of Thinking, Fast and Slow by Daniel Kahneman
System 1
“operates automatically and quickly, with little or no effort…”
System 2
“allocates attention to the effortful mental activities that demand it…”
Daniel Kahneman, 2011 Short excerpts from Chapter 1
Original book (2011) · Chapter 1 excerpt
Research example · see original source for evaluation conditions.

System 1 + System 2 Inside a VLA

GR00T N1 (NVIDIA, 2025)
GR00T N1 paper, Figure 2: original model overview showing the VLM as System 2 and diffusion transformer as System 1
GR00T N1 (NVIDIA, 2025), Figure 2 (cropped) · CC BY 4.0
Research example · see original source for evaluation conditions.

System 1 + System 2: Agent + VLA

Claude Code brand mark and wordmarkCodex — OpenAI brand markNVIDIA logo for GR00T VLA
System 2
System 1
Agent
Subtask
Observation
VLA
GR00T N1.6
Robot actions
Codex
Vesta paper (2026), Appendix B + D · Robot demos + harness
Research example · see original source for evaluation conditions.

FLUX 3 Action

Black Forest Labs logoFLUX 3 Action promotional bannerBlack Forest Labs: three block-stacking rollouts, FLUX 3 Action alone versus FLUX 3 Action plus Astra▶ Watch originalBFL reported robot-control benchmark results
Black Forest Labs · bfl.ai/models/flux-3-action
Research example · see original source for evaluation conditions.

ASPIRE: Success vs. Cost

10 LIBERO-Pro tasks · held-out success vs. estimated API cost
ASPIRE held-out success versus estimated total API cost, replotted from the authors’ Figure 8 and Table 18 data
ASPIRE Project · Jim Fan Post · Authors’ Chart Data
Research example · see original source for evaluation conditions.

ASPIRE: Learning on a Real Robot

Rubik’s Cube · tokens and time per completed face turn
ASPIRE real-robot Rubik’s Cube efficiency across three sequential episodes, with tokens per completed turn and minutes per completed turnASPIRE original example Rubik’s Cube turns video▶ Watch original
Example cube turns
ASPIRE Project · Jim Fan Post · Authors’ Chart Data
Research example · see original source for evaluation conditions.
Agents for Autonomous Research

ENPIRE Architecture

Agentic Robot Policy Self-Improvement in the Real World
ENPIRE Figure 2 architecture: environment construction, coding-agent harness, tool APIs, robot environment, and policy improvement
NVIDIA / CMU / UC Berkeley — Figure 2 · Project Page
Research example · see original source for evaluation conditions.

MRU: Mean Robot Utilization

ENPIRE: pin-insertion autoresearch with a robot fleet▶ Watch original
NVIDIA ENPIRE — Pin Insertion Fleet (8×) · Original Video
Research example · see original source for evaluation conditions.

System 3: Agentic Autoresearch

System 3
Systems 1 + 2
AutoresearchAgent
Propose + test
Rollouts + scores
Agent + VLA
Robot actions
Conceptual synthesis · ENPIRE paper · ENPIRE Project
Conceptual architecture · research direction, not a verified deployment.

Tatbot: A System 3 Architecture

NVIDIA server racks, representative cloud infrastructureNVIDIA DGX SparkNVIDIA Jetson AGX Thor developer kitClaude logo for Opus 5.5OpenAI logo for GPT-6.1 SolQwen official organization logoROS logoPython programming language logoRust programming language logoOpenCode coding-agent harness logoHugging Face logo for LeRobot
CLOUD MODELS
LOCAL · ON SITE
System 3
System 2
System 1
Server racks
DGX Spark
Jetson Thor
Opus 5.5
GPT-6.1 Sol
Qwen
Python
Rust
Improve
Execute
Logs + outcomes
LeRobot
DGX Spark · Jetson Thor · NVIDIA rack · Asset sources
Conceptual architecture · research direction, not a verified deployment.
Thank You
HICAMAustin Robotics