Files
lerobot/examples/rtc/README.md
T
Eugene Mironov 433ccc9603 Update README
2025-11-15 00:09:01 +07:00

252 lines
9.4 KiB
Markdown

# Real-Time Chunking (RTC) Examples
This directory contains examples and evaluation scripts for Real-Time Chunking (RTC), a technique for improving action chunking policies in real-time robot control.
## Overview
Real-Time Chunking addresses the challenge of maintaining consistency and reactivity when using action chunking policies with non-negligible inference latency. It uses a guidance technique during diffusion sampling to blend new action predictions with previously planned actions.
**Key Benefits:**
- Maintains consistency between consecutive action chunks
- Reduces jitter and improves smoothness
- Adapts to inference delays dynamically
**Reference:** [Physical Intelligence - Real-Time Chunking](https://www.physicalintelligence.company/download/real_time_chunking.pdf)
## Scripts
### 1. `eval_dataset.py`
Offline evaluation on dataset samples with detailed visualization and validation.
**Features:**
- Compare RTC vs non-RTC predictions on two random dataset samples
- Validate RTC behavior (delay region, blend region, post-horizon region)
- Generate debug visualizations:
- Denoising step comparisons (x_t, v_t, x1_t, corrections)
- Final action predictions comparison
- Support for torch.compile() optimization
- Memory-efficient sequential policy loading for large models
**Usage:**
```bash
# Basic usage with SmolVLA policy
uv run python examples/rtc/eval_dataset.py \
--policy.path=helper2424/smolvla_check_rtc_last3 \
--dataset.repo_id=helper2424/check_rtc \
--rtc.execution_horizon=8 \
--device=mps \
--rtc.max_guidance_weight=10.0 \
--seed=10
# With Pi0.5 policy on CUDA
uv run python examples/rtc/eval_dataset.py \
--policy.path=lerobot/pi05_libero_finetuned \
--dataset.repo_id=HuggingFaceVLA/libero \
--rtc.execution_horizon=8 \
--device=cuda
# With Pi0 policy
uv run python examples/rtc/eval_dataset.py \
--policy.path=lerobot/pi0_libero_finetuned \
--dataset.repo_id=HuggingFaceVLA/libero \
--rtc.execution_horizon=8 \
--device=cuda
# With torch.compile for faster inference
uv run python examples/rtc/eval_dataset.py \
--policy.path=helper2424/smolvla_check_rtc_last3 \
--dataset.repo_id=helper2424/check_rtc \
--rtc.execution_horizon=8 \
--device=cuda \
--use_torch_compile=true \
--torch_compile_mode=max-autotune
# Enable CUDA graphs (advanced - may cause tensor aliasing errors)
uv run python examples/rtc/eval_dataset.py \
--policy.path=helper2424/smolvla_check_rtc_last3 \
--dataset.repo_id=helper2424/check_rtc \
--use_torch_compile=true \
--torch_compile_backend=inductor \
--torch_compile_mode=max-autotune \
--torch_compile_disable_cudagraphs=false
```
**Key Parameters:**
- `--policy.path`: Path to pretrained policy
- `--dataset.repo_id`: Dataset to evaluate on
- `--rtc.execution_horizon`: Number of steps to maintain consistency (default: 20)
- `--rtc.max_guidance_weight`: Maximum guidance weight (default: 10.0)
- `--rtc.prefix_attention_schedule`: Schedule type (ZEROS, ONES, LINEAR, EXP)
- `--inference_delay`: Inference delay for RTC (default: 4)
- `--seed`: Random seed for reproducibility (default: 42)
- `--output_dir`: Directory to save visualizations (default: rtc_debug_output)
- `--device`: Device to use (cuda, cpu, mps, auto)
- `--use_torch_compile`: Enable torch.compile() for faster inference
**Output:**
The script generates several visualization files in `rtc_debug_output/`:
- `denoising_xt_comparison.png` - Noisy state evolution during denoising
- `denoising_vt_comparison.png` - Velocity predictions during denoising
- `denoising_x1t_comparison.png` - Predicted final states during denoising
- `denoising_correction_comparison.png` - RTC guidance corrections applied
- `final_actions_comparison.png` - Final action predictions (prev_chunk, no_rtc, rtc)
The script also validates RTC behavior and reports:
- ✅ Delay region [0:inference_delay]: RTC = prev_chunk
- ✅ Blend region [inference_delay:execution_horizon]: prev_chunk ≤ RTC ≤ no_rtc
- ✅ Post-horizon [execution_horizon:]: RTC = no_rtc
### 2. `eval_with_real_robot.py`
Real-time evaluation on physical robots or simulation environments.
**Features:**
- Run policy with RTC on real robot or simulation
- Multi-threaded action execution and inference
- Action queue management with proper timing
- Latency tracking and adaptive inference delay
- Support for both robots and gym environments
- Support for torch.compile() optimization
**Usage:**
```bash
# With real robot
uv run python examples/rtc/eval_with_real_robot.py \
--policy.path=lerobot/smolvla_base \
--robot.type=so100 \
--task="pick up the cup" \
--duration=30.0
# With simulation environment
uv run python examples/rtc/eval_with_real_robot.py \
--policy.path=lerobot/smolvla_base \
--env.type=pusht \
--duration=60.0
# With policy compilation (CUDA only, not MPS)
uv run python examples/rtc/eval_with_real_robot.py \
--policy.path=lerobot/smolvla_base \
--robot.type=so100 \
--use_torch_compile=true \
--torch_compile_mode=max-autotune
```
**Key Parameters:**
- `--policy.path`: Path to pretrained policy
- `--robot.type` or `--env.type`: Robot or environment to use
- `--task`: Task description (for VLA models)
- `--rtc.execution_horizon`: Number of steps to maintain consistency (default: 10)
- `--rtc.max_guidance_weight`: Maximum guidance weight (default: 1.0)
- `--rtc.prefix_attention_schedule`: Schedule type (ZEROS, ONES, LINEAR, EXP)
- `--duration`: How long to run (seconds, default: 30.0)
- `--fps`: Action execution frequency (Hz, default: 10.0)
- `--action_queue_size_to_get_new_actions`: Queue size threshold to request new actions (default: 30)
- `--device`: Device to use (cuda, cpu, mps, auto)
- `--use_torch_compile`: Enable torch.compile() for faster inference
## Understanding RTC Parameters
### `execution_horizon`
Number of timesteps from previous chunk to maintain consistency with. Higher values mean more consistency but potentially less reactivity.
**Typical values:** 8-12 steps for dataset evaluation, 10 steps for real-time execution
### `max_guidance_weight`
Upper bound on guidance strength. Higher values give stronger consistency but may over-constrain new predictions.
**Typical values:**
- Dataset evaluation: 10.0-100.0 (can be higher for analysis)
- Real-time execution: 1.0-10.0 (more conservative)
### `prefix_attention_schedule`
How to weight consistency across the overlap region:
- `ZEROS`: Binary (full weight up to inference_delay, then zero)
- `ONES`: Full weight across entire execution_horizon
- `LINEAR`: Linear decay from inference_delay to execution_horizon
- `EXP`: Exponential decay (recommended)
**Recommended:** `EXP`
### `inference_delay`
Number of timesteps from the prefix to use for guidance. Typically calculated dynamically based on inference latency in real-time execution, but fixed for dataset evaluation.
**Typical values:** 3-5 steps for dataset evaluation
### `action_queue_size_to_get_new_actions` (real-time only)
Threshold for requesting new action chunks. Should be higher than `inference_delay + execution_horizon` to ensure smooth operation.
**Typical values:** 20-30 steps
## Validation Rules (Dataset Evaluation)
The dataset evaluation script validates that RTC behavior matches expectations:
1. **Delay Region [0:inference_delay]**: RTC actions should equal previous chunk
- Ensures consistency during the inference delay period
2. **Blend Region [inference_delay:execution_horizon]**: RTC should be between prev_chunk and no_rtc
- Smooth transition from previous plan to new predictions
3. **Post-Horizon [execution_horizon:]**: RTC should equal no_rtc
- Full adoption of new predictions after execution horizon
## Tips
1. **Start with dataset evaluation** (`eval_dataset.py`) to understand RTC behavior and tune parameters before running on robot
2. **Use visualizations** to debug unexpected behavior - check denoising steps and final actions
3. **Tune execution_horizon** based on your inference latency and action frequency
4. **Monitor validation output** - failures indicate potential implementation issues or misconfigured parameters
5. **Compare different schedules** - EXP usually works best but LINEAR can be more interpretable
## Troubleshooting
### Validation fails in delay region
- Check that `prev_chunk_left_over` is properly passed to the policy
- Verify RTC guidance is being applied during denoising
- Look at denoising visualizations to see where guidance diverges
### Validation fails in post-horizon region
- RTC and no_rtc use different noise - verify same noise is being used for comparison
- Check that weights are correctly zeroed out after execution horizon
- Review prefix_attention_schedule visualization
### Poor performance on real robot
- Increase `action_queue_size_to_get_new_actions` if you see warnings
- Reduce `max_guidance_weight` if robot is too conservative
- Try different `prefix_attention_schedule` values
- Enable torch.compile() for faster inference (CUDA only)
### Memory issues with large models
- The dataset evaluation script loads policies sequentially to minimize memory
- For real-time execution, only one policy is loaded
- Use smaller batch sizes if needed
## Related Documentation
- [RTC Implementation](../../src/lerobot/policies/rtc/modeling_rtc.py)
- [RTC Configuration](../../src/lerobot/policies/rtc/configuration_rtc.py)
- [Action Queue](../../src/lerobot/policies/rtc/action_queue.py)
- [Physical Intelligence Paper](https://www.physicalintelligence.company/download/real_time_chunking.pdf)