AI

GTA-2 Builds Robot Skills From Axes, Not End-to-End Actions

Shar Hendrix 4 min read

End-to-end VLAs hide the plan. Code-as-Policies dumps a whole program in one shot. GTA-2 (Grounded Task Axes v2), from Carnegie Mellon’s Robotics Institute and Bosch Research, splits the difference: four vision-language agents write an inspectable skill, then a UR5e runs it.

The paper (arXiv:2609.09808, submitted September 9, 2026) reports 73.9% average zero-shot success on 14 real manipulation tasks, 31.4 points above the strongest baseline. After targeted human feedback on a failed stage, that average rises to 90.7%. No task demos, no policy training, no fine-tuning.

UR5e erasing a red scribble and pouring from a mug, with GTA-2 controller overlays
Erase and pour as composed controllers, then re-grounded in new scenes. Source: GTA-2 paper, Figure 1.

Four agents, one controller library

Hardware is a UR5e, a Robotiq 2F-85 gripper, and a calibrated ZED 2i. All four agents use Gemini 3.1 Pro, differing only in prompts and structured outputs:

  1. Task Decomposer — ordered subtasks from the instruction.
  2. Skill Generator — keypoints, axes, and controller recipes per subtask.
  3. Parameter Setter — numbers: offsets, forces, angles.
  4. Vision Module — metric 3D features from RGB-D.

Controllers include position and axis alignment, waypoints, constant force, gripper state, and a learned spiral DMP. Lower-priority commands sit in the null space of higher ones. If the pour height is wrong, you poke the Parameter Setter and keep the rest.

GTA-2 four-agent pipeline from task prompt to grounded robot script
Decompose, compose, set parameters, ground in RGB-D. Source: GTA-2 project page.
Fourteen UR5e tabletop tasks including microwave, dustpan, ironing, and pouring
The 14-task suite, from ball-in-basket to iron-the-tie. Source: GTA-2 paper, Figure 3.

Numbers against π0.5 and two code baselines

Each generalization cell is 20 physical trials. Zero-shot GTA-2 averages 73.9%. CaP-Primitive (Code-as-Policies with Cartesian motion, gripper, and contact force) is the best baseline at 42.5%. π0.5 (DROID-finetuned, run on a Franka for that baseline) and CaP-TAC (same task-axis library, one generated script) trail further. CaP-TAC does beat primitives on pouring, cutting, and sweeping, where contact geometry matters.

Plan generation is a separate study: 15 runs per task, 210 total. Zero-shot execution succeeds in 154/210 (73.3%). One feedback round reaches 181/210 (86.2%), two rounds 208/210 (99.0%), all 210 within three. First-round edits mostly hit the Skill Generator; later edits shift to vision grounding.

The authors are clear about the box they built. Features are grounded once, not tracked online. Slip or a moved object needs a new cycle. The controller library has to already express the objective. Tests are one tabletop arm.

A Human’s Take

I smiled at “fix the typo” rearranging letters into TEACH. That is a system that followed the words, not a hidden demo script. The useful part is the edit loop: you do not regenerate a 400-line policy because the spiral radius was 2 cm too small. I still want to see this on a second arm and a moving object. Until the vision module updates while the cup is pouring, it is a very smart first take, not a shift worker.

Sources