flow-robot-distillation: checkpoints
Parameter checkpoints for the single-training-seed pilot in https://github.com/akashprakas/flow-robot-distillation, which consolidates two single-task RL specialists (Robomimic MH Square, PH Tool Hang) into one multi-task flow-matching policy.
All results are from one training seed. Evaluation uses 64 fixed seeds (10000..10063) per task.
Files
| File | Arm | Description |
|---|---|---|
a1_shared_bc_230k.pkl |
A1 | Shared multi-task BC base, step 230k |
a2_square_actor.pkl, a2_toolhang_actor.pkl |
A2 | Specialist actors, step 1M |
a2_square_teacher.pkl, a2_toolhang_teacher.pkl |
A2 | Specialist teachers (actor + critics), step 1M |
a3_Q1.pkl, a3_Q2.pkl, a3_Q3.pkl |
A3 | Offline distillation students, step 20k, budgets Q1-Q3 |
a4_Q1.pkl, a4_Q2.pkl, a4_Q3.pkl |
A4 | Routed DAgger students, step 20k, budgets Q1-Q3 |
SHA256SUMS lists the checksums. Verify with sha256sum -c SHA256SUMS.
Security note
The files are Python pickles of JAX/NumPy parameter trees. Unpickling can execute arbitrary code, so verify the checksums and only load files from this repository.
Usage
git clone --recurse-submodules https://github.com/akashprakas/flow-robot-distillation.git
cd flow-robot-distillation && pip install -e .
hf download akashp805/flow-robot-distillation-checkpoints --local-dir checkpoints
python scripts/06_eval_64seeds.py \
--checkpoint checkpoints/a4_Q3.pkl \
--task square --modes ode sde \
--n_eval_envs 8 --eval_episodes 8
Expected A4 Q3 results (64 seeds): Square 87.50% ODE / 87.50% SDE, Tool Hang 73.44% ODE / 76.56% SDE.
The saved summaries are in the GitHub repository under results/data/.
Caveats
- Single training seed; no multi-seed confirmation.
- Results depend on the library versions and hardware (GPU vs CPU A4 runs differ).
- Tool Hang requires raw (not standardised) observations; see the repository README.