CaP-X Agentic Robotics

LLM agents that write Python code to control real robots — from zero-shot manipulation to RL-trained coding agents with sim-to-real transfer.

Based on CaP-X (NVIDIA, Berkeley, Stanford, CMU). arXiv: 2603.22435. MIT License.

Quick Start

1. Clone and Install

git clone https://github.com/capgym/cap-x.git
cd cap-x

# Requires Python 3.10 (BEHAVIOR) or 3.12 (RL/LIBERO)
uv sync          # uses uv (Astral) for package management

2. Start Perception Services

CaP-X runs perception models as microservices:

# SAM3 segmentation (port 8114)
python -m capx.perception.sam3_server

# Molmo 2 pointing (port 8117)
python -m capx.perception.molmo_server

# ContactGraspNet 6-DOF grasps (port 8115)
python -m capx.perception.grasp_server

# OWL-ViT detection (port 8118)
python -m capx.perception.owlvit_server

Requires CUDA GPU. For IK solvers:

PyRoKi (port 8116) — CPU-friendly inverse kinematics
cuRobo — GPU-accelerated motion planning (NVIDIA, requires CUDA)

3. Run a Benchmark Task

# Single task evaluation (zero-shot, 100 trials)
python scripts/eval.py \
  --task cube_lift \
  --model openai/gpt-5.2 \
  --tier S1 \
  --num_trials 100

# Full CaP-Bench sweep
python scripts/eval_capbench.py --model anthropic/claude-opus-4.5

Architecture

CaP-X
├── CaP-Gym ─── Gymnasium interface wrapping 187 tasks
│   ├── RoboSuite (7 core) ── Franka Panda tabletop/bimanual
│   ├── LIBERO-PRO (130+) ── Franka Panda kitchen/living
│   └── BEHAVIOR (50) ────── R1Pro humanoid, Isaac Sim
│
├── CaP-Bench ── 8 tiers (S1-S4 single, M1-M4 multi-turn)
│   ├── Varies: perception noise, API abstraction, visual grounding
│   └── 12 frontier LLMs/VLMs benchmarked
│
├── CaP-Agent0 ── Training-free agentic harness
│   ├── Visual Differencing Module (VDM)
│   ├── Auto-synthesized skill library (Voyager lineage)
│   └── Parallel ensembled reasoning (multi-model)
│
└── CaP-RL ──── GRPO post-training on code generation
    ├── 7B model: 25% → 80% in 50 iterations
    └── Zero-shot sim-to-real transfer

Control Flow

The agent never outputs raw joint commands. Instead:

LLM → generates Python code → composes perception + control APIs → robot executes

API abstraction levels (Franka example):

FrankaControlApi — Full high-level (perception + IK control)
FrankaControlPrivilegedApi — Oracle state (no perception noise)
FrankaControlReducedApi — Low-level primitives
FrankaControlReducedSkillLibraryApi — Low-level + auto-synthesized skills

CaP-Bench Tiers

Tier	Mode	Perception	Abstraction	Visual Grounding
S1	Single	Noiseless	High	--
S2	Single	Noisy	High	--
S3	Single	Noisy	Low	--
S4	Single	Noisy	Low	--
M1	Multi	Noisy	High	--
M2	Multi	Noisy	High	Multimodal feedback
M3	Multi	Noisy	Low	--
M4	Multi	Noisy	Low	VDM

Run specific tiers:

python scripts/eval.py --task cube_stack --model google/gemini-3-pro --tier M4

CaP-Agent0: Training-Free Harness

Three components inspired by Voyager (Wang et al., 2023):

1. Visual Differencing Module (VDM)

Convert before/after scene images into structured text describing what changed. Solves cross-modal alignment failures found in M2 tier.

2. Auto-Synthesized Skill Library

Reusable functions discovered from successful execution traces. Persist across trials. Compilation pipeline:

# After running evaluations, compile skill library from successful traces
python scripts/skill_library_compilation/compile.py \
  --eval_outputs results/cube_lift/ \
  --output skills/

# Use compiled library in evaluation
python scripts/eval.py --task cube_stack --tier M4 --skill_library skills/

9 task-agnostic skills discovered (geometric utilities, grasp filters, quaternion helpers).

3. Parallel Ensembled Reasoning

Multiple models generate candidate solutions:

python scripts/eval_agent0.py \
  --task cube_stack \
  --models "google/gemini-3-pro,openai/gpt-5.2,anthropic/claude-opus-4.5" \
  --ensemble

Result: Matches/exceeds human expert code on 4/7 tasks. Competitive with trained VLA policies (OpenVLA, pi_0) despite being training-free.

CaP-RL: RL on Code Generation

GRPO (Group Relative Policy Optimization) applied to Qwen2.5-Coder-7B:

# Train on privileged tier S1 for stable convergence
python scripts/train_rl.py \
  --task cube_lift \
  --base_model Qwen/Qwen2.5-Coder-7B-Instruct \
  --group_size 15 \
  --train_iterations 50 \
  --gpu_type h100

Task	Base (7B)	+CaP-RL (50 iter)	Human Expert
Cube Lift (sim)	25%	80%	93%
Cube Stack (sim)	4%	44%	73%
Spill Wipe (sim)	30%	93%	100%
Cube Lift (real)	24%	84%	92%
Cube Stack (real)	12%	76%	84%

Transfer zero-shot to real robots because reasoning is over abstract APIs, not raw pixels.

See references/caprl-training.md for full GRPO configuration, GPU requirements, and training recipes.

Sim-to-Real Transfer

Simulation (CaP-Gym) ──[abstract APIs]──> Real Robot
                                          ├── Franka Panda (primary)
                                          ├── AgiBot G1 (bimanual demos)
                                          └── R1Pro (mobile manipulation)

Real-world requirements:

Stereo camera with calibrated metric-scale depth (e.g., ZED)
robots_realtime package for Franka Panda control
Same perception service stack (SAM3, Molmo, ContactGraspNet)

Extending CaP-X

Add a New Task

# Register in capx/envs/your_suite/
class MyTask(CaPGymTask):
    """Gymnasium-compatible task wrapper."""

    def __init__(self):
        self.perception = PerceptionStack(sam3=True, molmo=True, depth=True)
        self.control = FrankaControlApi(ik_solver="pyroki")

    def get_prompt(self) -> str:
        return "Pick up the red cube and place it on the green platform."

    def compute_reward(self, obs, action, next_obs) -> float:
        # Verifiable environment reward for RLVR
        return float(self._check_cube_on_platform(next_obs))

Add a New Robot

Implement the control API interface. See references/api-spec.md for the full specification:

goto_pose(pos, quat) — Move end-effector to target pose via IK
grasp() / open_gripper() / close_gripper() — Gripper control
get_ee_pose() — Current end-effector state

Add a Perception Module

Input: RGB image (+ optional depth)
Output: Structured detection/segmentation result
Follow the pattern in capx/perception/ servers

Integration with Life Agent OS

CaP-X connects to the Broomva Agent OS stack at three points:

Arcan Orchestration

Orchestrate multi-step robotic workflows as Arcan task graphs:

// Arcan pipeline for robotic manipulation
let pipeline = arcan::Pipeline::new()
    .step("perceive", capx_perception_step)   // SAM3 + Molmo
    .step("plan", capx_agent_step)            // LLM code generation
    .step("execute", capx_control_step)       // Robot actuation
    .step("verify", capx_vdm_step)            // Visual differencing
    .on_failure("reflect", capx_reflect_step) // Self-correction loop
    .build();

Spaces Agent Networking

Robotic agents publish events to Spaces channels:

#robot-logs    <- Execution traces, success/failure, skill library updates
#agent-logs    <- Session summaries (standard bstack integration)
#perception    <- Scene descriptions, object detections, grasp candidates

Multiple robot agents share skill libraries via Spaces — a skill discovered by one robot transfers to others.

Lago Persistence

Store and version:

Skill libraries — SHA-256 hashed Python functions as Lago blobs
Evaluation traces — Execution logs for CaP-RL training data
Benchmark results — CaP-Bench scores across models and tiers
Trained checkpoints — RL-tuned model weights

Companion Skills

# ORCA Hand (17-DOF tendon-driven hand — ETH Zurich)
npx skills add broomva/skills --skill orcahand -g -y

# Remote GPU (offload perception/training to NUC or cloud)
npx skills add broomva/skills --skill remote-gpu -g -y

Resources

references/

references/api-spec.md — Full perception and control API specification
references/capbench-results.md — Benchmark results across 12 models, 8 tiers
references/caprl-training.md — CaP-RL training guide (GRPO config, GPU requirements, recipes)

scripts/

scripts/setup-perception.sh — Start all perception microservices
scripts/run-benchmark.sh — Full CaP-Bench evaluation sweep

Key Papers

CaP-X: arXiv:2603.22435 — Framework paper
Voyager: arXiv:2305.16291 — Skill library + self-reflection origin
Code as Policies: arXiv:2209.07753 — LLM code to robot control
SayCan: arXiv:2204.01691 — Grounding LLMs in robotic affordances

capx-agentic-robotics

SKILL.mdskills/robotics/capx-agentic-robotics/SKILL.md